"Farshid Dabaghi Thesis" - Institut Camille Jordan

N◦ d’ordre: 2014–ISAL–0070
Thèse de Doctorat
Spécialité
Mathématiques Appliquées
Présentée par
Farshid Dabaghi
Pour obtenir le grade de
Docteur de l’Institut National des Sciences Appliquées de Lyon
Sujet de thèse:
Étude de la convergence des méthodes de redistribution de
masse pour les problèmes de contact en élastodynamique
Soutenue le 8 juillet 2014 devant le jury composé de:
Mikaël Barboteu
Université de Perpignan
Examinateur
Marius Cocou
Université de Aix–Marseille
Rapporteur
Pavel Krejčí
Académie des Sciences de la République Tchèque
Examinateur
Manuel D. P. Monteiro Marques
Université de Lisbonne, Portugal
Rapporteur
Laetitia Paoli
Université de Saint–Etienne
Examinateur
Adrien Petrov
INSA de Lyon
Directeur
Jérôme Pousin
INSA de Lyon
Directeur
Yves Renard
INSA de Lyon
Directeur
For the guardians of the kurdish people, the beauty of our lands.
For the courageous mountains of Kurdistan
Pour les gardiens du peuple kurde, la beauté de nos terres.
Pour les courageuses montagnes du Kurdistan
Je dédie cette thèse à la mémoire de mon grand–père dont j’aurais tellement aimé qu’il soit
présent mais la nature en a décidé autrement.
Je souhaite dédier en particulier cette thèse à mon père, ma mère et tous les autres membres de ma famille. Malgré la distance qui nous sépare, je n’ai jamais manqué de soutien et
d’encouragements de leurs part.
Remerciement
Je tiens à remercier tout particulièrement mes directeurs de thèse Adrien Petrov, Jérôme
Pousin et Yves Renard pour toute l’aide qu’ils m’ont apporté et pour leur patience.
Je tiens ici à exprimer mes sentiments respectueux à Marius Cocou, Professeur des universités
au Laboratoire de Mécanique et d’Acoustique CNRS, et Manuel D. P. Monteiro Marques, Professeur à l’université de Lisbonne, d’avoir accepté d’être rapporteurs ainsi que Mikaël Barboteu,
Maître de Conférences à l’université de Perpignan, Pavel Krejčí, chercheur à l?Institut de Mathmatiques de l’Académie des Sciences de la République Tchèque, et Laetitia Paoli, Professeur à
l’université de Saint–Étienne, de participer au jury.
Je garde un bon souvenir des moments passés avec tous les membres du Pôle de Mathématiques de l’INSA de Lyon et du laboratoire de l’Institut Camille Jordan.
Je remercie profondément les doctorants de l’Institut Camille Jordan et les invités tunisiens
avec qui j’ai passé des moments inoubliables: Saber Amdouni, Éric Dalissier, Mathieu Fabre,
Amin Karoui, Khadidja Ben mansour, Imen Mekkaoui, Rafik Ben Hassine et Rabii Mlika.
Et enfin un grand merci à ma famille, à mes proches et à mes camarades proches ou lointains
en particulier Behzad Nasri, Ayoub Moradi et Farah Laariedh, pour leurs encouragements, leurs
soutiens et pour leurs amitiés.
iii
Contents
Présentation de la thèse
Introduction . . . . . . . . . .
Un peu d’histoire . . . . . . .
Etat de l’art . . . . . . . . . .
Formulations mathématiques
Résumé des travaux . . . . .
Conclusion et perspectives . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
1
2
4
7
9
20
1 Convergence of mass redistribution method for the one–dimensional wave
equation with a unilateral constraint at the boundary
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Existence and uniqueness results by using the characteristic method . . . . . . .
1.3 The equivalence between the variational formulation and the differential inclusion
1.4 Convergence of mass redistribution method . . . . . . . . . . . . . . . . . . . . .
1.5 Numerical examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Appendix A . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Appendix B . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Appendix C . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
23
23
25
29
33
39
43
46
47
2 Numerical approximations of a one dimensional elastodynamic contact
lem based on mass redistribution method
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Description of the models . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3 Discretization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3.1 Classical finite element discretization . . . . . . . . . . . . . . . . . .
2.3.2 Finite element discretization with the mass redistribution method . .
2.4 The wave equation with Signorini and Neumann boundary conditions . . . .
2.4.1 Analytical solution . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.4.2 Numerical simulations . . . . . . . . . . . . . . . . . . . . . . . . . .
2.5 The wave equation with Signorini and Dirichlet boundary conditions . . . .
2.5.1 Analytical solution . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.5.2 Numerical simulations . . . . . . . . . . . . . . . . . . . . . . . . . .
Appendix D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Appendix E . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
55
56
57
62
62
63
65
65
66
69
69
70
82
84
prob.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
3 Efficiency of a weighted mass redistribution method for an elastodynamic
93
problem with unilateral constraints
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
3.2 Finite element discretization and convergence of the mass redistributions method 95
3.3 The wave equation with Signorini and Dirichlet boundary conditions . . . . . . . 101
Contents
3.3.1
3.3.2
Analytical solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
Comparisons between different mass redistributions for some time–space
discretisations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
4 Numerical study of convergence of the mass redistribution method
dynamic contact problems
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2 The finite element approximation of problem . . . . . . . . . . . . . .
4.2.1 The semi–discretization in space . . . . . . . . . . . . . . . . .
4.2.2 The time integration methods . . . . . . . . . . . . . . . . . . .
4.3 Numerical experiments . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3.1 The Crank–Nicolson scheme . . . . . . . . . . . . . . . . . . . .
4.3.2 The backward Euler method . . . . . . . . . . . . . . . . . . . .
4.4 Some complements . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.4.1 The Newmark method I . . . . . . . . . . . . . . . . . . . . . .
4.4.2 The Newmark method II . . . . . . . . . . . . . . . . . . . . . .
4.4.3 The convergence rates . . . . . . . . . . . . . . . . . . . . . . .
Bibliography
for elasto109
. . . . . . 109
. . . . . . 111
. . . . . . 111
. . . . . . 112
. . . . . . 113
. . . . . . 114
. . . . . . 120
. . . . . . 123
. . . . . . 123
. . . . . . 127
. . . . . . 130
135
vi
Présentation de la thèse
Sommaire
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1
Un peu d’histoire . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2
Etat de l’art . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4
Formulations mathématiques . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7
Résumé des travaux . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9
Conclusion et perspectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
20
Introduction
Un ou plusieurs corps déformables en mouvement peuvent entrer en contact en un point, le
long d’une ligne ou d’une surface. Les systèmes aux dérivées partielles sont particulièrement
bien adaptés pour modéliser des mouvements complexes de corps soumis à des conditions aux
limites faisant intervenir les conditions de contact. Ces problèmes de contact occupent un rôle
prépondérant dans de nombreux domaines comme par exemple dans l’industrie automobile où
le contact entre le pneu et la chaussée (cf. Figure 1) ainsi que les crash tests (cf. Figure 2)
intéressent particulièrement les ingénieurs ou encore en génie civil où les ouvrages d’arts de plus
en plus complexes tels que les ponts ou encore les gratte ciel nécessitent une étude préalable des
matériaux en contact qui vont devoir subir d’importants efforts.
Figure 1: La simulation de roulement du pneu par Brinkmeier et al [1]
La modélisation des phénomènes de contact pose de sérieuses difficultés du point de vue
mathématique. Plus précisément, suite à l’application d’une charge sur deux corps en contact,
la surface de contact devient inconnue a priori. Ainsi, la modélisation des problèmes de contact
conduit à des équations non–linéaires équivalentes à un système d’inéquations à faible régularité
Présentation de la thèse
de la solution. En présence de frottement et dans le cas multi–corps, le problème de contact
devient plus compliqué. La formulation d’un problème bien posé présente un défi majeur pour
le théoricien surtout dans le cas de dimensions supérieures à un et de problèmes complexes
(frottement, multi–corps...). D’un point de vue numérique, l’établissement d’un schéma robuste
et efficace reste encore un challenge important.
Figure 2: simulation d’accident de voiture par le groupe de "FEA–Opt Technology"
(http://www.fea-optimization.com)
Un peu d’histoire
Certaines découvertes archéologiques faites en Egypte montrent que les bâtisseurs des pyramides
s’intéressaient déjà aux problèmes de contact et de frottement. Ainsi des textes expliquant partiellement certaines techniques employées pour transporter des pierres de plusieurs tonnes sur
des centaines de métres ont été mis à jour récemment par des archéologues. Par ailleurs, il existe
également un certain nombre de bas reliefs comme celui trouvé à Ninive (cf. Figure 3) représentant le transport de divers objets de grande taille et mettant en évidence que ces problèmes
intéressent l’homme depuis la nuit des temps.
Les premiers travaux significatifs dans le domaine du contact et du frottement ont été effectués
par L. De Vinci au XVIème siècle. Il a établi les deux lois fondamentales de frottement; la
première indique que le frottement est proportionnel à la force normale et la seconde met en
évidence que les zones de contact n’ont pas d’effet sur le frottement. G. Amontons a redécouvert
les deux lois fondamentales de frottement au XVII siècle qui n’ont pas été publiées par Léonard
de Vinci. Il faut attendre le XVIIIème siècle et L. Euler qui en dérivant l’angle au repos d’une
masse sur un plan incliné fait la distinction entre les frottements statique et cinétique. L. Euler
écrivit dans [2]: "L’expérience nous ayant fait voir que la force du frottement est toujours égale
à une certaine partie de la pression, dont un corps est pressé contre la surface, sur laquelle il se
meut, de sorte que le frottement ne dépend ni de la grandeur de la base, dont le corps touche
la surface, ni du degré de vitesse, il n’est pas diffcile de déterminer l’effet du frottement dans
toutes sortes de machines par le moyen du calcul: vu que le frottement doit être regardé comme
une force constante, qui est toujours directement contraire à la direction du mouvement, et
qui agit dans une direction, qui passe par le plan de l’attouchement des corps qui se meuvent
2
l’un sur l’autre". Ch.–A. de Coulomb a formulé mathématiquement les lois de Amontons dans
la deuxième moitié du XVIIIème siécle. Il a ainsi déterminé l’ampleur du frottement lors de
mouvements de glissement et il a étudié également l’influence de quatre facteurs principaux sur
le frottement: la nature des matériaux et leurs revêtements de surface, l’étendue de la surface,
la pression normale et la durée de temps pendant laquelle les surfaces sont restées en contact
(temps de repos). De plus, il a pris en considération dans ses travaux l’influence de la vitesse de
glissement, de la température et de l’humidité [3].
Figure 3: Une statue étant transporté sur un traîneau, Ninive, environ 700 avant J.–C., British
Museum
Figure 4: Les expériences de Coulomb
Au XIXème siècle, H. R. Hertz dans [4] s’est intéressé à des surfaces sans frottement et des solides
élastiques linéaires non–conformes. Cependant, seulement quelques problèmes de contact peuvent
être résolus analytiquement, H. R. Hertz s’est essentiellement penché sur des formulations simples
et à des modéles idéalistes qui ne sont pas suffisamment réguliers pour aboutir à une solution
3
Présentation de la thèse
dans les cas complexes. Néanmoins des contributions dans cette direction ont été obtenues dans
[5, 6, 7, 8, 9]. Enfin, A. Signorini a introduit une formulation mathématique pour le problème
général de l’équilibre d’un corps élastique en contact sans frottement avec une structure rigide
[10]. G. Fichera a appelé les conditions de contact introduites par A. Signorini, les conditions
de Signorini. Il s’est par ailleurs intéressé à l’existence de l’équilibre statique par minimisation
de l’énergie totale sur un convexe dans [11]. En particulier, il est le premier à avoir formulé le
problème de contact en utilisant les inéquations variationnelles. Cette formulation mathématique
du contact est à l’origine de nombreux travaux utilisant différentes théories comme les inéquations
variationnelles, l’analyse de Fourier, la théorie du point fixe ou encore la méthode des opérateurs
d’hystérésis.
Etat de l’art
Les travaux mathématiques s’articulent autour de la formulation de la condition de contact
introduite par A. Signorini. Cette condition peut être relaxée et remplacée par un problème
approché faisant intervenir une pénalisation appelée encore compliance normale (cf. [12, 13, 14,
15, 16, 17, 18, 19, 20, 21]). Le problème ainsi obtenu est beaucoup plus simple à étudier, il est
ainsi aisé de prouver l’existence et l’unicité d’une solution en utilisant différentes théories comme
les inéquations variationnelles ou encore la méthode du point fixe. Ensuite, toute la difficulté
réside dans le passage à la limite dans le problème approché qui s’obtient généralement en
effectuant des estimations a priori. Il existe par ailleurs d’autres approches permettant d’obtenir
l’existence ainsi que l’unicité pour des problèmes de contact comme par exemple la reformulation
du problème en utilisant les inclusions différentielles ou encore la discrétisation. Nous ferons
dans la suite une synthèse des résultats existants pour les problèmes de contact statiques et
dynamiques ainsi que des méthodes numériques permettant d’obtenir une solution approchée
pour ces problèmes.
Le premier résultat d’existence d’une solution pour un problème statique modélisant un
corps élastique en contact avec un obstacle rigide sans frottement est obtenu dans [11]. Quant
au problème avec frottement de Tresca, il a été étudié quelques années plus tard dans [22]. Le
problème de Signorini avec frottement de Coulomb a été traité dans [23]. Plus précisément, sous
l’hypothèse que le coefficient de frottement est suffisamment petit, l’existence d’une solution a
été obtenue en utilisant l’algorithme itératif de Panagiotopoulos [24] ainsi que le théorème du
point fixe. Par ailleurs, les méthodes de "shift" (cf. [25, 26]) ont permis de mettre en évidence
un résultat de régularité. Certaines techniques employées dans [23] ont été utilisées ensuite pour
traiter des domaines coercifs [27] et semi–coercifs [28]. Ces résultats ont été redémontrés par des
méthodes de pénalisation dans [17, 18]. Remarquons que l’existence de solutions pour le problème
de Signorini statique avec une régularisation du terme de frottement a été prouvée dans [29]. En
ce qui concerne la non–unicité de la solution de ce problème avec un coefficient de frottement
suffisamment grand a été mis en évidence dans [30, 31]. D’autres contre–exemples d’unicité de
solution ont été présentés récemment pour le cas statique et quasi–statique dans [32] et dans [33],
respectivement. Par ailleurs, dans [34] un exemple de non–existence pour une problème quasi–
statique avec frottement est étudié. Enfin, sous des hypothèses sur le déplacement tangentiel et
pour un coefficient de frottement suffisamment petit, un résultat d’unicité a été démontré dans
4
[35]. Le lecteur est également invité à consulter [36] où un certain nombre de résultats sur les
problèmes de contact statiques et quasi–statiques sont présentés.
L’étude mathématique de problèmes de contact dynamiques est généralement plus délicate
que celle des problèmes de contact statiques ou quasi–statiques. Les difficultés mathématiques
sont liées à la nature hyperbolique des problèmes élastodynamiques combinée à la contrainte exprimée en déplacement conduisant à des irréguliarités intrinsèques du problème. Notons que dans
certains travaux (cf. [22]), la contrainte en déplacement est remplacée par la contrainte en vitesse
levant ainsi une difficulté majeure à l’étude de problèmes de contact dynamiques. Commençons
par nous intéresser aux problèmes de contact dynamiques monodimensionnel. Sous l’hypothèse
de choc élastique parfait, l’existence et l’unicité pour un problème de corde vibrante avec obstacle
rigide et un obstacle ponctuel et avec une loi d’impact arbitraire sont établis respectivement dans
[37] et [38]. Ensuite, des résultats d’existence et d’unicité ont été obtenus pour une corde vibrante
avec un obstacle continu dans [39], avec un obstacle concave dans [40], avec un obstacle ponctuel
dans [41] et avec obstacle adhésif dans [42]. Enfin le mouvement de deux cordes vibrantes se
rejoignant en un point est traité dans [43]. Quant aux problèmes de contact dynamiques en
dimension supérieure à 1, il existe très peu de résultats significatifs dans la litérature. Ainsi
l’existence et l’unicité d’une solution est établie pour l’équation des ondes avec des contraintes
unilatérales dans un demi–espace en utilisant des techniques microlocales dans [44], il n’a pas
été cependant possible d’étendre ces techniques à domaines bornés. L’existence de la solution
pour un domaine borné utilisant une technique de compacité par compensation est étudiée dans
[45]. L’existence de la solution est etablie pour un problème de contact multidimensionnelle, en
dynamique et quasi–statique, avec prise en compte d’un terme thermo–élastique dans [46]. La
question de l’unicité et de la conservation de l’énergie restent des problèmes ouverts en dimension
supérieure à 1. Certaines difficultés rencontrées pour des problèmes de contact elastodynamiques
peuvent être levées en ajoutant de la viscosité dans les équations (cf. [47, 48, 49, 50, 51, 52]). Par
ailleurs, les espaces fonctionnels caractérisant les traces ont été identifiés dans [53] pour l’équation
des ondes amorties ainsi que pour l’evolution d’un matériau viscoélastique de Kelvin–Voigt, les
deux problèmes ont des conditions unilatérales aux bords. Cependant ces résultats de régularité
ne permettent pas de déduire que les pertes d’énergie sont purement visqueuses. Il est toutefois possible sous certaines hypothèses supplémentaires de construire une solution permettant de
conclure que les pertes d’énergie sont purement visqueuses (cf. [54, 55]).
Comme nous l’avons évoqué précédemment, l’unicité d’un problème de contact reste ouvert
en dimension strictement supérieure à un et pour un domaine quelconque. C’est pour cette raison
que les approches numériques pour résoudre les problèmes complexes sont utilisés. En effet, ces
approches ne cessent de se développer grâce à l’evolution rapide des capacités des ordinateurs.
En dépit de ce développement, la modélisation correcte des problèmes de contact reste encore
un challenge à cause de la non–linéarité du contact et du frottement .
La méthode des élèments finis est largement utilisée pour simuler numériquement des problèmes de contact et consiste à discrétiser la géométrie du corps en un ensemble d’éléments de
taille finie. La première application de la méthode des éléments finis pour les problèmes de contact a été faite dans [56, 57, 58]. Dans ces travaux, les auteurs se sont intéressés aux problèmes
d’élasticité sans frottement. Ils ont été suivis par plusieurs autres travaux, citons notamment
5
Présentation de la thèse
[59, 60]. A noter que la première analyse mathématique de convergence a été effectuée dans
[61]. La solution du problème de contact avec frottement par éléments finis est traitée dans [62].
Dans [63], les auteurs ont développé une procédure utilisée dans [60] permettant de tenir compte
des effets de frottement. D’autres analyses par éléments finis ont été réalisées afin de décrire le
problème de contact avec frottement en élastostatique dans [64] et en élastodynamique dans [65].
Imposer la condition de contact dans le cadre de la méthode des élements finis (autrement
dit transformer les inégalités variationnelles en égalités) est un autre sujet important pour les
aspects numériques. Les deux méthodes les plus utilisées sont les multiplicateurs de Lagrange et
la méthode de pénalisation. La premiére méthode a été introduite dans [59], et elle consiste à
écrire la contrainte de contact en utilisant la force de contact. Une méthode pour inclure les effets
de frottement en utilisant des facteurs de frottement de Coulomb en termes de multiplicateurs de
Lagrange est introduite dans [66]. Les premiers travaux a avoir utilisé la méthode de pénalisation
pour les problèmes de contact sont [57, 67, 62, 68]. Dans cette méthode, l’interpénétration de
deux surfaces de contact est possible et la force de contact normale est liée à la pénétration
par un paramètre arbitraire de pénalisation. Pour plus d’informations sur les avantages et les
inconvénients des méthodes, de multiplicateurs de Lagrange et de pénalisation, voir [69]. On
peut citer aussi les autres techniques qui existent dans la littérature, comme la méthode de la
flexibilité [60], l’approche de programmation mathématique [70, 71], la méthode du Lagrangien
augmenté [72]. La méthode des éléments finis classiques combine une discrétisation en espace
et un schéma d’intégration en temps pour résoudre le problème semi–discret. Dans le cadre
de problèmes linéaires, un schéma converge s’il vérifie la stabilité linéaire (par exemple, zéro–
stabilité et A–stabilité [73]) et la condition de consistance, lorsque la taille du pas de temps tend
vers zéro. Dans les problèmes élastodynamiques linéaires simples, on peut utiliser un schéma
explicite conditionnellement stable. Pour les cas complexes (par exemple pour les matériaux
incompressibles), les schémas implicites inconditionnellement stables ont été développés dans
[74, 75]. Ces schémas perdent leurs stabilités inconditionnelles dans les cas des problèmes non–
linéaires. Pour ce type de problèmes, d’autres critères de stabilité doivent être utilisés, comme
par exemple, l’analyse énergétique du problème. Il est donc très difficile de construire un schéma
numérique stable pour les problèmes dynamiques avec contraintes unilatérales. Par conséquent,
deux critères importants pour la stabilité d’une méthode sont utilisées pour justifier les simulations numériques pour les problèmes de contact. Ces critères sont le comportement de l’énergie
et les oscillations parasites du déplacement et des forces de contact [76]. Après nombreux efforts,
une méthode efficace et robuste a été proposée. Une possibilité consiste à ajouter des termes
d’amortissement mais nous avons une perte de la précision de la solution. Une autre possibilité
est d’impliciter la force de contact [77] mais cela conduit à une perte d’énergie, même avec un
raffinement du pas de temps. Certaines méthodes d’intégration temporelle sont proposées dans
les articles [78, 79, 80, 81], elles conservent l’énergie mais n’évitent pas complètement des oscillations du déplacement et de la force de contact. Une autre façon d’éviter les oscillations parasites
est d’imposer la compatibilité de la vitesse et de l’accélération des points de contact grâce aux
vecteurs d’erreur correspondants (voir par exemple [82, 83]). Cette méthode n’est pas totalement
satisfaisante. Il est bien connu que la semi–discrétisation à l’aide des éléments finis classiques
conduit à un problème mal posé qui est provoqué par la multiplicité de solution du problème [84].
L’ajout d’une loi d’impact supplémentaire, permet d’obtenir un problème semi–discret bien posé
6
[85], mais cela conduit à des oscillations parasites. Une autre approche consistant à supprimer
la masse au niveau du nœud de contact est proposée dans [84] appelée la méthode de redistribution de masse, cette méthode évite les oscillations parasites au bord de contact. Le problème
semi–discret, en utilisant la méthode de redistribution de masse, est un problème bien posé sans
oscillations parasites et la solution conserve l’énergie (voir [84] et chapitre 1). La convergence du
problème semi–discret pour des matériaux visco–élastiques pour la dimension deux et trois est
établie dans [86]. Nous avons établi la convergence de la méthode des matériaux élastiques en
dimension un (voir chapitre 1).
Formulations mathématiques
La condition de contact introduite par A. Signorini s’écrit
(1)
uν ≤ 0,
rν ≤ 0,
uν rν = 0,
où uν et rν sont respectivement le déplacement normal et la force de contact. La condition (1)
indique que si le contact n’est pas établi, uν < 0 alors rν = 0, autrement dit la réaction entre le
matériau et l’obstacle rigide n’existe pas. Lorsque le contact est établi, (1) permet de déduire
que uν = 0 et rν < 0 en d’autres termes une réaction existe entre le matériau et l’obstacle rigide.
def
def ∂u
Nous utiliserons dans suite les notations ut = ∂u
∂t et ux = ∂x .
On considère un domaine borné Ωd ⊂ Rd avec d ≤ 3 ainsi que sa frontière
Dir
Neu
∂Ωd = Γd ∪ Γd
Sig
∪ Γd
Neu et ΓSig représentent respectivement les frontières correspondantes aux conditions
où ΓDir
d , Γd
d
de Dirichlet, Neumann et au contact unilatéral. On définit le déplacement u(x, t) pour (x, t) ∈
Ωd × [0, T ]. Le problème mathématique qui nous intéresse s’écrit de la manière suivante:
(2)
ρutt − div σ(u) = ` dans Ωd × (0, T ),
où ρ est la densité du matériau considéré et σ(u) est le tenseur de contraintes vérifiant
σij (u) = Aijkl εkl (u) dans Ωd ,
def
où Aijkl est le tenseur d’élasticité et le tenseur des déformations linéarisé défini par
εkl (u) =
1 ( ∂uk ∂ul )
+
.
2 ∂xl ∂xk
Nous supposerons que le matériau est homogène et isotrope. En utilisant la loi de Hooke, le
tenseur d’élasticité s’exprime alors en fonction des constantes de Lamé λ et µ comme suit
Aijkl = λδij δkl + 2µδkl δjk .
def
Par conséquent, le tenseur des contraintes peut se réécrire de la manière suivante:
σ(u) = λTr(ε)I + 2µε,
7
Présentation de la thèse
où δij , Tr et I sont respectivement le symbole de Kronecker, la trace et le tenseur identité. Le
problème (2) est complété par des conditions initiales données par
u(·, 0) = u0
(3)
et ut (·, 0) = v 0
dans Ωd ,
et des conditions aux limites
(4a)
u = α sur ΓDir
d × (0, T ) et σ(u)ν = g
sur ΓNeu
× (0, T ),
d
(4b)
0 ≥ ν T u ⊥ ν T (σ(u)ν) ≤ 0 et τi T (σ(u)ν) = 0,
i = 1, 2 sur ΓSig
d × (0, T ),
où ν et τ désignent la normale unitaire extérieure et la tangente unitaire, α est un vecteur
donné et g est une force de surfacique dépendant de l’espace, (·)T est la transposée de (·)
L’orthogonalité a ici le sens suivant; si nous avons suffisamment de régularité, cela signifie que le
produit ν T u(ν T σ(u)ν) est nul presque partout sur la frontière. Si ce n’est pas le cas, l’inégalité
(4b) est intégrée sur un ensemble approprié de fonctions de test, ce qui conduit à une formulation
faible pour la condition unilatérale. En particulier, ΓDir
d = ∅ dans le cas où nous considérons le
problème d’une barre en chute libre heurtant un obstacle rigide.
Nous allons maintenant définir la formulation faible associée au problème (2)–(4). Nous introduisons au préalable le convexe
K = {u ∈ U : u(·, t) ∈ V p.p. t et ν T u ≤ 0 p.p. sur ΓSig
d },
def
2
2
2
où V = {u ∈ H1 (Ωd ) : u = 0 p.p. sur ΓDir
d } et U = {u ∈ L (0, T ; V) : ut ∈ L (0, T ; L (Ωd ))}
d
avec X désignant l’espace X . Nous obtenons la formulation faible associée à (2)–(4) en multipliant (2) par v−u, v ∈ K, puis en intégrant formellement cette expression sur Ωd , on obtient:


trouver u : [0, T ] → K∫ telle que

∫
∫

∫
g·(v − u) dx,
`·(v − u) dx +
Aijkl εij (u)εkl (v − u) dx ≥
ρutt ·(v−u) dx +

ΓNeu
Ωd
Ωd
Ωd

d


pour tout v ∈ K.
def
def
Dans les trois premiers chapitres de cette thèse, nous allons nous restreindre au cas monodimensionel qui après adimensionnement s’écrit comme suit:
utt − uxx = ` dans Ω1 × (0, T ),
(5)
avec les conditions initiales
u(·, 0) = u0
(6)
et ut (·, 0) = v 0
sur Ω1 ,
et des conditions aux limites
(7)
u(L, ·) = α,
ux (L, ·) = g
et 0 ≤ u(0, ·) ⊥ ux (0, ·) ≤ 0 sur [0, T ].
Comme précédemment, nous allons définir la formulation faible associée au problème (5)–(7).
Soit le convexe
def
K = {u ∈ U : u(·, t) ∈ V p.p. t et u(0) ≥ 0},
8
où V = {u ∈ H1 (0, L) : u(L) = 0} et U = {u ∈ L2 (0, T ; V) : ut ∈ L2 (0, T ; L2 (0, L))}. Par
conséquent, la formulation faible associée à (2)–(4) est obtenue en multipliant (2) par v − u,
v ∈ K, puis en intégrant cette expression formellement sur (0, L) × (0, T ), ce qui nous donne


u ∈ K tel que
trouver

∫ L
∫ T∫ L



0
0

v (v(·, 0)−u ) dx −
ut (vt −ut ) dx dt
−
0 ∫ 0
∫ 0T ∫ L
(8)
T



+
ux (vx −ux ) dx dt ≥
`(v−u),



0
0
0


pour tous v ∈ K pour lesquels il existe ζ > 0 avec v = u pour t ≥ T − ζ.
def
def
Résumé des travaux
Cette thèse s’insére dans la problèmatique d’une bonne approximation numérique des problèmes
de contact en élastodynamique. En particulier, nous nous attachons à caractériser les propriétés
de la méthode de redistribution de masse (stabilité et convergence) qui consiste à redistribuer la
masse des nœuds de contact sur les autres nœuds. La thèse est organisée en quatre chapitres.
Les trois premiers chapitres sont consacrés à l’étude du mouvement d’une barre pouvant heurter
un obstacle rigide. La méthode de redistribution de masse est introduite et son efficacité est
validée numériquement. Le dernier chapitre de cette thèse est dédié au problème en dimension 2
et 3. Toutes les simulations numériques présentées ont été effectuées en utilisant la bibliothèque
d’éléments finis Getfem++ (cf. [87]).
Chapitre 1: Convergence de la méthode de redistribution de masse pour
l’équation des ondes monodimensionnelle avec une contrainte unilatérale à
la frontière.
Nous nous intéressons dans ce chapitre au mouvement d’une barre élastique de longueur L
dont l’une des extrémités est encastrée et l’autre est libre de bouger tant qu’elle ne touche pas
un obstacle rigide (cf. Figure 5). On considère le déplacement vertical u(x, t) de cette barre
L
x=0
Figure 5: Mouvement d’une barre élastique
qui dépend du point matériel x et de l’instant t, et satisfaisant (5)–(7) avec `(x, t) = 0 pour
tout (x, t) ∈ [0, L] × [0, T ]. Nous allons maintenant réécrire le problème (5)–(7) sous forme d’un
problème d’inclusion différentielle en utilisant la méthode des caractéristiques. Plus précisément,
9
Présentation de la thèse
def
def
en introduisant le changement de variables ξ = x + t et η = x − t, il est possible de décomposer
la solution u en la somme de deux fonctions p et q de la manière suivante:
u(ξ, η) = p(ξ) + q(η),
ce qui conduit à l’expression
u(x, t) = p(x+t) + q(x−t).
Les conditions initiales ainsi que la condition de contact peuvent être réécrites comme suit
p(x) + q(x) = u0 (x) et p0 (x) − q 0 (x) = v 0 (x),
0 ≤ p(t) + q(−t) ⊥ p0 (t) + q 0 (−t) ≤ 0 et p(L+t) + q(L−t) = 0.
La sous–différentielle JN de la fonction indicatrice qui est une application multivoque définie par


si x < 0,

{0}
def
JN [x] = [0, +∞) si x = 0,


∅
si x > 0,
permet de reformuler le problème (5)–(7) et d’écrire
(9a)
f 0 (t) ∈ −JN [f ](t) − 2p0 (t) p.p. t ∈ (0, L),
(9b)
f (0) = −u0 (0).
Les propriétés de monotonie et de convexité de l’opérateur conduisent au théorème d’existence
et d’unicité pour le problème (9) explicité ci–dessous.
Théorème 1. Supposons que p(t) ∈ W1,1 (0, L). Alors le problème de Cauchy (9) admet une
solution unique f (t) ∈ W1,1 (0, L).
Nous allons maintenant établir que (8) et (9) sont équivalents. Pour cela nous divisons le domaine
[0, L] × [0, T ] en quatre domaines I–IV comme représenté sur la Figure 6.
t
(L, L)
(0, L)
II
I
III
IV
x
(L, 0)
Figure 6: Les quatre domaines permettant de déterminer u
Dans chacun des domaines, une expression de la solution u en fonction de p et q est obtenue
conduisant au théorème suivant:
10
Théorème 2. Supposons que u0 and v 0 appartiennnent respectivement à K et L2 (0, L). Alors
f est solution du problème (9) si et seulement si u est une solution de (8) pour T = L
Par ailleurs, la solution u conserve l’énergie, c’est–à–dire nous avons
∫
∫
1 L 00
1 L 00
E(t) =
|u (x)+v 0 (x)|2 dx +
|u (x)−v 0 (x)|2 dx
4 0
4 0
est constante.
Nous allons introduire maintenant des problèmes discretisés en temps et en espace approchant
la formulation faible (8). Dans un premier temps, un problème semi–discretisé en espace est
introduit. Ce problème utilise la méthode des éléments finis affines de Lagrange associé à une
méthode des multiplicateurs de Lagrange. Plus précisément, la solution u est approchée par
uh (x, t) =
n−1
∑
uj (t)ϕj (x),
j=0
i|
où la fonction de base ϕi (x) est definie comme suit ϕi (x) = 1− |x−a
lorsque x ∈ [amax(i−1,0) , ai+1 ]
h
def
def
et ϕi (x) = 0 ailleurs. Il en découle que la formulation faible (8) est approchée par


trouver uh : [0, T ] → Vh et λ : [0, T ] → R tels que vh ∈ Vh

∫ L
∫ L




uh,x vh,x = −λvh (0) p.p. t ∈ [0, T ],
uh,tt vh dx +
(10)
0
0



0 ≤ uh (0, ·) ⊥ λ ≤ 0 p.p. t ∈ [0, T ],



u (·, 0) = u0 et u (·, 0) = v 0 .
h
h
h,t
h
Notons que (10) peut être réécrit de la manière suivante:


trouver U : [0, T ] → Rn et λ : [0, T ] → R tels que



M U + SU = −λe
p.p. t ∈ [0, T ],
tt
0
(11)

0 ≤ u0 ⊥ λ ≤ 0 p.p. t ∈ [0, T ],




U (0) = U 0 et Ut (0) = V 0 ,
def
où M et S sont appelées respectivement les matrices de masse et de rigidité et e0 = (1, 0, . . . , 0)T .
Nous pouvons remarquer que (11) n’est a priori pas bien posé car l’équation au nœud de contact
peut avoir plusieurs solutions. La méthode de redistribution de masse est alors introduite. Cette
méthode est basée sur une redistribution de masse de manière à ce qu’il n’y ait pas d’inertie
au niveau du nœud de contact, supprimant ainsi l’opérateur différentiel en temps au nœud de
contact. En d’autres termes, la matrice de masse M est remplacée par la matrice suivante
(
)
0 0
def
M mod =
,
¯
0 M
¯ ij = Mi+1,j+1 pour tous i, j = 1, . . . , n − 2. Il est alors commode d’introduire les notations
où M
suivantes:
(
)
(
)
S00 C T
u0
S=
et U =
,
¯
C
S¯
U
11
Présentation de la thèse
def
def ∫
¯ def
où U
= (u1 , . . . , un−1 )T et S¯ij = Si+1,j+1 avec C = Ω ϕ0i+1 ϕ00 dx, i = 0, . . . , n − 2. Avec ces
notations et en se servant de la méthode de redistribution de masse, le problème se réécrit comme
suit
(12)


trouver U : [0, T ] → Rn et λ : [0, T ] → R tels que




¯¯
¯¯


 M Utt + S U = −Cu0 p.p. t ∈ [0, T ],
¯ = −λ p.p. t ∈ [0, T ],
S00 u0 + C T U


 0 ≤ u ⊥ λ ≤ 0 p.p. t ∈ [0, T ],

0



 U (0) = U 0 et U (0) = V 0 .
t
¯ n’est pas singulière, on déduit que (12) possède une unique solution (U, λ)
Puisque la matrice M
qui est lipschitzienne. Par ailleurs, cette solution conserve l’énergie. En choisissant de manière
judicieuse un projecteur qui conserve la contrainte au point de contact puis en effectuant des
estimations a priori, le théorème suivant en découle:
Théorème 3. Supposons que limh→0 (kuh (·, 0)−u0 kH1 (0,L) + kuh,t (·, 0)−v0 kL2 (0,L) ) = 0. Alors la
solution du problème approché (12) converge vers la solution unique de la formulation faible (8).
Enfin la dernière partie du premier chapitre est consacrée à quelques tests numériques qui confirment l’efficacité de la méthode de redistribution de masse. En effet, nous observons sur la Figure
7 (gauche) que plus les pas de temps ∆t et d’espace ∆x sont raffinés plus la solution approchée est
proche de la solution exacte. Nous pouvons faire la même observation pour l’évolution d’énergie
(cf. Figure 7 (droite)).
Figure 7: Convergence numérique de la solution associée au problème avec la méthode de
redistribution de masse vers la solution exacte au point de contact x = 0 (à gauche). Convergence
numérique de l’évolution de l’énergie (à droite).
12
Chapitre 2: Approximations numériques d’un problème de contact élastodynamique unidimensionnel basé sur la méthode de redistribution de masse.
Nous comparons ici différents schémas numériques que nous avons développés pour le problème
introduit dans le chapitre précédent avec cette fois–ci un second membre non nul `(x, t), dépendant de x et de t. Nous analysons également le cas de la chute libre d’une barre qui heurte
un obstacle rigide. Pour ces deux problèmes, une solution explicite est mise en évidence puis
une comparaison est effectuée avec les solutions approcheés obtenues pour différents schémas
numériques utilisant notamment la méthode de redistribution de masse.
Nous commençons par établir que le problème (5)–(7) possède une solution unique. Pour
cela, nous introduisons u
¯ qui est la solution unique du problème auxiliaire (5)–(7) où la condition
def
contact a été remplacée par une condition Dirichlet homogène. Ensuite, nous définissons v = u−¯
u
comme la solution du problème
vtt (x, t) − vxx (x, t) = 0,
(x, t) ∈ (0, L) × (0, T ),
v(x, 0) = 0 et vt (x, 0) = 0,
x ∈ (0, L),
0 ≤ v(0, t) ⊥ vx (0, t) + u
¯x (0, t) ≤ 0 et v(L, t) = 0,
t ∈ [0, T ].
def
En utilisant la méthode des caractéristiques ainsi que la notation w(t) = −p(t) − q(−t),
nous sommes en mesure de montrer que le problème précédent est équivalent à un problème d’inclusion différentielle. Ce problème est ensuite régularisé puis des estimations a priori sont obtenues permettant de conclure que notre problème admet une unique solution
u ∈ L∞ (0, T ; V) ∩ W1,∞ (0, T ; H) sous des hypothèses de régularité sur les données initiales
ainsi que sur la force externe `.
La stratégie adoptée pour obtenir le problème semi–discrétisé en espace est la même que celle
présentée dans le premier chapitre; la seule différence vient du second membre qui est ici différent
de zéro. Plus précisément, la méthode des éléments finis affines de Lagrange classique combinée
avec la matrice de masse modifiée ou la matrice de masse non modifiée (appelée matrice de masse
standard dans la suite) est étudiée. Des méthodes d’intégration en temps permettant d’éviter
des oscillations artificielles pour le déplacement et pour la force de contact conservant l’énergie
sont proposées. Nous avons ainsi utilisé trois méthodes d’intégration temporelle.
Les méthodes de Newmark:


trouver Uhn+1 : [0, T ] → Rn et λn+1 : [0, T ] → R tel que:

(
)



n+1
n + ∆tU n + 1 −β ∆t2 U n + β∆t2 U n+1 ,

U
=
U

h,t
h
h,tt
h,tt
2
 h
n+1
n + (1−γ)∆tU n + γ∆tU n+1 ,
Uh,t
= Uh,t
h,tt
h,tt



n+1
n+1
n+1 e + F n+1 ,

M
U
+
SU
=
−λ

0
h,tt
h



n+1 ≤ 0.
0 ≤ un+1
⊥
λ
0
13
Présentation de la thèse
La θ–méthode:


trouver Uhn+1 : [0, T ] → Rn et λn+1 : [0, T ] → R tel que:




U n+1 = U n + ∆t((1−θ)U n +θU n+1 ),

h
h,t
h,t
 h
n+1
n + ∆t((1−θ)U n +θU n+1 ),
Uh,t
= Uh,t
h,tt
h,tt


n+1
n+1

n+1
M Uh,tt + SUh = −λ
e0 + F n+1 ,



0 ≤ un+1 ⊥ λn+1 ≤ 0.
0
La méthode de Paoli–Schatzman:


trouver Uhn+1 : [0, T ] → Rn et λn : [0, T ] → R tel que:





M (Uhn+1 −2Uhn +Uhn−1 )


+ S(βUhn+1 +(1−2β)Uhn +βUhn−1 ) = −λn e0 pour tous n ≥ 2,
2
∆tn+1
n−1

0 ≤ un,e = u0 + eu0

⊥ λn ≤ 0,

0

1
+
e



U0 et U1 donné.
Chaque méthode proposée ci–dessus est testée avec une matrice de masse standard ainsi qu’une
matrice de masse modifiée puis elle est comparée aux solutions analytiques. Notons que la
solution analytique obtenue pour la chute libre d’une barre n’est pas suffisamment régulière, ce
qui ne permet pas de mettre en évidence l’efficacité de la méthode de redistribution de masse.
Concernant le second problème traité dans ce chapitre, une solution explicite périodique a été
mise en évidence. Plus précisément, nous avons déterminé la solution en utilisant la méthode
des caractéristiques sur trois phases du phénomène; avant, durant et après le contact. Dans
les simulations numériques présentées ci–dessous, nous avons calculé la solution approchée en
utilisant la méthode de Crank–Nicolson. Tout d’abord, nous avons comparé trois solutions
approchées avec la solution analytique, en utilisant la matrice de masse standard et en choisissant
des pas d’espace ∆x et de temps ∆t de telle sorte qu’ils tendent vers zéro (pour les solutions
approchées). Puis ensuite, nous avons effectué un travail analogue en remplaçant la matrice
de masse standard par la matrice de masse modifée. Enfin nous avons comparé les résultats
obtenus pour les deux matrices de masse pour le déplacement de tous les points de la barre.
Le même travail a été effectué pour la force de contact ainsi que pour l’énergie. Nous avons
également vérifié la convergence vers la solution analytique (cf. Figure 8). Nous présentons ci–
dessus des courbes de convergence pour le déplacement dans les normes Lp (0, T ; L2 (0, L)) et
Lp (0, T ; H1 (0, L)) avec p = 2, ∞, pour la force de contact et pour le multiplicateur de Lagrange
dans la norme L2 (0, T ) et pour l’énergie dans les normes Lp (0, T ) avec p = 2, ∞, pour les
méthodes de masse standard et modifée (cf. Figure 9). Nous avons mené les mêmes investigations
pour des méthodes de Newmark en prenant comme paramètres (β, γ) = ( 21 , 21 ) et (β, γ) = ( 12 , 1),
pour la méthode d’Euler implicite et pour la méthode Paoli–Schatzman avec des coefficients de
restitution e = 0, 21 , 1.
Les tests numériques ont montré que la méthode de redistribution de masse élimine les oscillations du déplacement, réduit les oscillations des forces de contact et conserve l’énergie à de
très petites oscillations près. De plus la méthode de redistribution de masse améliore les taux de
convergence.
14
Figure 8: Comparaison de la solution analytique (u, λ) et des solutions approchées (Uhn , λn )
obtenues en utilisant les matrices de masse standard (à gauche) et modifiée (à droite) au nœud
de contact.
Chapitre 3 : Sur l’efficacité d’une méthode de redistribution de masse
pondérée pour un problème élastodynamique avec des contraintes unilatérales.
Nous considérons, comme dans le chapitre 1, le mouvement d’une barre élastique dont l’une de
ses extrémités est encastrée et l’autre est libre de tout mouvement tant qu’elle ne touche pas un
obstacle rigide. La formulation mathématique de ce problème est donnée par (5)–(7). Ensuite
le problème semi–discretisé en espace est introduit en utilisant encore une fois la méthode des
éléments finis affines de Lagrange associée à un multiplicateur de Lagrange. La différence fondamentale avec les travaux précédents réside dans le fait que la masse enlevée au nœud de contact
est maintenant redistribuée sur les autres nœuds et non plus juste supprimée (cf. chapitres 1 et 2).
Par conséquent, la masse totale du matériau est préservée. Par ailleurs, une fonction pondérée
permettant de redistribuer la masse du nœud de contact sur les autres nœuds tout en éliminant
la masse du nœud de contact
est introduite. La matrice de masse modifiée M mod est main∫
def
L
tenant définie par Mijmod = 0 ϕi ϕj wh,ij dx tel que wh,0i = wh,i0 pour tout i = 0, . . . , n − 1 et
wh,ij = 1 + o(h) for all i, j = 1, . . . , n − 1 où wh est la fonction pondérée mentionnée plus haut.
15
Présentation de la thèse
(a) kuh −ukL∞ (0,T ;H)
(b) kuh −ukL2 (0,T ;H)
Figure 9: Comparaison des courbes d’erreurs obtenues avec les matrices de masse standard et
modifiées.
En utilisant les notations du chapitre 1 et en se servant de la méthode de redistribution de masse,
la formulation faible (8) est approchée par

n


trouver U : [0, T ] → R et λ : [0, T ] → R such that


¯¯
¯¯
¯


M Utt + S U = −Cu0 + F p.p. t ∈ [0, T ],
¯ = −λ + F0 a.e. t ∈ [0, T ],
(13)
S00 u0 + C T U




0 ≤ u0 ⊥ λ ≤ 0 a.e. t ∈ [0, T ],



U (0) = U 0 et U (0) = V 0 ,
t
∫L
∫L
def
où F¯ = ( 0 f ϕ1 dx, . . . , 0 f ϕn−1 dx)T . Ensuite une estimation d’erreur en temps pour la
méthode de Crank–Nicolson a été obtenue conduisant au théorème suivant:
T
Théorème 4. Soit N = ∆t
∈ N pour T > 0 donné. Alors, l’erreur de troncature pour la méthode
de Crank–Nicolson permettant de résoudre le problème semi–discret (13) est d’ordre ∆t.
Il est possible ensuite de montrer que (13) est équivalent à


trouver uh : [0, T ] → Vh telle que pour tout vh ∈ K ∩ Vh


∫ T∫ L






 0 0 uh,tt (x, t)(vh (x, t)−uh (x, t))wh (x) dx dt
∫ L∫ T
(14)

+
uh,x (x, t)(vh,x (x, t)−uh,x (x, t)) dx dt



0 ∫0

∫

L
T



f (x, t)(vh (x, t)−uh (x, t)) dx dt.
≥
0
0
Le résultat de convergence est résumé dans le théorème suivant:
Théorème 5. Supposons que limh→0 (kuh (·, 0)−u0 kH1 (0,L) + kuh,t (·, 0)−v0 kL2 (0,L) ) = 0. Alors la
solution du problème approché (14) converge vers l’unique solution de la formulation faible (8)
lorsque h tend vers 0.
16
Dans les simulations numériques, nous avons considéré trois répartitions de masses correspondants respectivement à la non répartition de la masse du nœud de contact (Mod 1), à la répartition de la masse du nœud de contact uniformément sur les autres nœuds (Mod 2) ainsi qu’à la
répartition de la masse du nœud de contact sur le nœud précédant le nœud de contact (Mod 3)
(cf. Figure 10). La solution approchée est d’autant meilleure que la répartition de la masse du
nœud de contact est faite sur un autre nœud proche du nœud de contact. En particulier, la
meilleure approximation est obtenue dans le cas où toute la répartition de masse est faite sur
le nœud qui précéde le contact (cf. Figure 10 (Mod 3)). Pour les simulations numériques, nous
avons utilisé trois méthode d’intégration en temps; les méthodes de Crank–Nicolson, d’Euler
implicite et de Paoli–Schatzman dans le cas particulier où (β, e) = ( 14 , 0).
Exact solution
Mod 1
Mod 2
Mod 3
0.6
0.5
Displacement in the contact point
Displacement in the contact point
0.7
0.4
0.3
0.2
0.1
0
Exact solution
Mod 1
Mod 2
Mod 3
0.4
0.3
0.2
0.1
0
−0.1
−0.2
0
1
2
3
Time
4
5
−0.1
0
6
0.2
0.4
0.6
Time
0.8
1
1.2
0.6
0.4
Lagrange multiplier (λ)
Lagrange multiplier (λ)
0.5
Exact solution
Mod 1
Mod 2
Mod 3
0
−0.5
−1
0.2
Exact solution
Mod 1
Mod 2
Mod 3
0
−0.2
−0.4
−0.6
−0.8
−1.5
0
1
2
3
Time
4
5
−1
0.8
6
0.85
0.9
0.95
Time
1
1.05
1.1
Figure 10: Comparaison de la solution analytique (u, λ) et des solutions approchées (Uhn , λn ) par
des matrices de masse modifiées dans le nœud de contact.
17
Présentation de la thèse
Chapitre 4 : Etude numérique de la convergence de la méthode de redistribution de masse pour les problèmes de contact élastodynamique
L’objectif du chapitre 4 est l’étude numérique de la convergence de la méthode de redistribution
de masse pour des problèmes de contact élastodynamique en dimension deux et trois. Comme
dans les chapitres précédents, nous considérons le contact d’une barre élastique et d’un obstacle
rigide dont l’une des extrémités de la barre est encastrée et l’autre peut se déplacer librement
tant qu’elle ne touche pas un obstacle rigide. La formulation mathématique de ce problème est
donnée par (2)–(4). Ensuite, nous avons utilisé la méthode des éléments finis pour obtenir le
problème semi–discrétisé ainsi que la matrice de masse modifiée. Par conséquent, le problème
approché de (2)–(4) s’écrit

N ⊥ : [0, T ] → N ⊥ et λ : [0, T ] → Rnc telle que

trouver UN

h : [0, T ] → N , Uh


Mmod UN ⊥ + S(UN + U N ⊥ ) = F + BT λ pour p.p. t ∈ [0, T ],
h,tt
h
h
T UN ⊥ λ ≤ 0 pour tous i ∈ I et pour p.p. t ∈ [0, T ],

0
≥
ν

i
c
i
h



0
0
Uh (0) = Uh et Uh,t (0) = Vh .
Normal displacemnt on contact boundary
Normal displacemnt on contact boundary
Les méthodes d’intégration en temps de Newmark ainsi que de Euler implicite ont été utilisées
pour résoudre le problème semi–discret. Comme dans le second chapitre, chaque schéma est testé
avec les matrices de masse standard et modifiée. En dimension deux, nous pouvons comparer
le déplacement normal au niveau des nœuds de contact, la force de contact et l’évolution de
l’énergie pour les deux matrices de masse (cf. Figures 11 et 12).
0.8
0.6
0.4
0.2
0
−0.2
0.1
0.08
0.06
0.04
0.02
Space
0 0
2
4
6
8
10
0.8
0.6
0.4
0.2
0
−0.2
0.1
0.08
10
0.06
8
6
0.04
4
0.02
Space
Time
18
2
0 0
Time
Figure 11: Solutions approchées obtenues en utilisant la méthode de Crank–Nicolson avec des
matrices de masse standard (gauche) et modifiées (droite) aux nœuds de contact.
Figure 12: Energie associée à la solution approchée pour les matrices de masse standard (gauche)
et modifiées (droite) avec le schéma de Crank–Nicolson.
En raison de la croissance rapide du nombre de degrés de liberté en dimension trois, nous avons
pu obtenir des courbes de convergence avec seulement trois points. Cependant les courbes de
convergence confirment l’efficacité de la matrice de masse modifiée.
Contrairement au problème monodimensionnel, nous n’avons pas pu déterminer une solution
analytique associée à (2)–(4). Toutefois une solution de référence peut être obtenue en prenant
∆t et ∆x très petits. Nous présentons des courbes de convergence, en dimension deux, pour le
déplacement, la force de contact ainsi que l’énergie.
19
Présentation de la thèse
(a) kUn
h −UkL∞ (0,T ;H)
(b) kUn
h −UkL2 (0,T ;H)
Figure 13: Comparaison des courbes de convergence obtenues en utilisant les matrices de masse
standard et modifiée avec des éléments carrés et ∆x
∆t = 2 pour le cas 2D.
(a) kUn
h −UkL∞ (0,T ;H)
(b) kUn
h −UkL2 (0,T ;H)
Figure 14: Comparaison des courbes de convergence obtenues en utilisant les matrices de masse
standard et modifiée avec des éléments cubiques et ∆x
∆t = 2 pour le cas 3D.
Conclusion et perspectives
Les principales contributions de ce travail de thèse sont l’étude théorique et numérique de la
méthode de redistribution de masse en élastodynamique avec impact. Une attention particulière
a été portée à l’étude numérique des taux de convergence. Plus précisément, nous avons mis
en évidence que la méthode de redistribution de masse était particulièrement bien adaptée pour
approcher des problèmes elastodynamiques avec des contraintes unilatérales. En effet, cette
méthode élimine les oscillations parasites, elle conserve l’énergie et elle améliore également les
taux de convergence. Par ailleurs, nous avons établi que malgré la faible régularité des problèmes
d’impact, les schémas d’ordre 2 en temps donnent des taux de convergence meilleurs. Enfin nous
20
avons mis en évidence que la solution approchée est d’autant meilleure que la répartition de
masse du nœud de contact est distribuée sur un nœud proche du nœud de contact. Nous avons
validé ce résultat pour des problèmes 1D. Puis nous pouvons remarquer que cette constatation
reste valable en dimension supérieure et en particulier en 2D comme l’indique le Tableau 1. En
Method
kUhn −U kL∞ (0,T ;H)
Std
0.0104
Mod 1
0.0046
Mod 3
0.0038
Table 1: Estimation d’erreur avec ∆x1 = ∆x2 = 0.05
effet, la répartition de masse sur le nœud précédant le contact (Mod 3) donne une estimation
d’erreur bien meilleure que dans le cas où la matrice standard est utilisée (Std) ou encore dans
le cas où la masse est simplement éliminée sur les nœuds qui sont en contact (Mod 2).
Les perspectives sont nombreuses à court et à long terme. Il serait intéressant de voir si
les méthodes numériques développées dans cette thèse donnent des résultats similaires pour
d’autres matériaux comme par exemple les matériaux à mémoire de forme ou encore les matériaux
ferromagnétiques. Par ailleurs, les estimations d’erreur en temps et en espace représentent un
challenge important. En effet, dans le cas monodimensionnel le problème continu est bien posé,
une estimation d’erreur de la discrétisation devrait être possible à obtenir. La principale difficulté
consiste à déterminer un lien entre la formulation éléments finis et la formulation en terme
d’inclusion différentielle. Enfin, il semble que l’introduction d’une méthode à poids devrait
permettre une écriture unique du problème approché conduisant ensuite à une estimation d’erreur
en espace.
21
Chapter 1
Convergence of mass redistribution
method for the one–dimensional wave
equation with a unilateral constraint at
the boundary
This chapter is published in ESAIM: Mathematical Modelling and Numerical Analysis [88]
Contents
1.1
1.1
Introduction
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
23
1.2
Existence and uniqueness results by using the characteristic method .
25
1.3
The equivalence between the variational formulation and the differential inclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
29
1.4
Convergence of mass redistribution method . . . . . . . . . . . . . . . .
33
1.5
Numerical examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
39
Appendix A . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
43
Appendix B . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
46
Appendix C . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
47
Introduction
We consider an elastic bar of length L vibrating vertically. More precisely, one end of the bar is
free to move, as long as it does not hit a material obstacle, while the other end is clamped (see
Fig. 1.1). The obstacle constrains the displacement of the extremity to be greater than or equal
to 0.
We describe now the mathematical situation. We assume that the material of the bar is
homogeneous and satisfies the theory of small deformations. Let x be the spatial coordinate
along the bar, with the origin at the material obstacle, let u(x, t) be the displacement at time
t ∈ [0, T ], T > 0 , of the material point of spatial coordinate x ∈ [0, L]. Then the mathematical
problem can be formulated as follows:
(1.1)
u
¨(x, t) − u00 (x, t) = 0,
(x, t) ∈ (0, L) × (0, T ),
Chapter 1. Convergence of mass redistribution method for the one–dimensional
wave equation with a unilateral constraint at the boundary
with Cauchy initial data
u(x, 0) = u0 (x) and u(x,
˙
0) = v 0 (x),
(1.2)
x ∈ (0, L),
and Signorini and Dirichlet boundary conditions at x = 0 and x = L, respectively,
0 ≤ u(0, t) ⊥ u0 (0, t) ≤ 0 and u(L, t) = 0,
(1.3)
def
t ∈ [0, T ],
def
∂u
0
where u˙ = ∂u
∂t and u = ∂x . The orthogonality has the natural meaning; namely if we have
enough regularity, it means that at the boundary x = 0 the product u(0, ·)u0 (0, ·) vanishes for
almost every t. If it is not the case, the above inequality is integrated on an appropriate set of
test functions, leading to a weak formulation for the unilateral condition.
L
x=0
Figure 1.1: An elastic bar vibrating on impacting obstacle.
Observe that from mathematical viewpoint, the Signorini conditions mean that when the bar
touches the obstacle in x = 0, its reaction can be only upwards, so that u0 (0, t) ≤ 0 on the set
{t : u(0, t) = 0}. While in the case where the bar does not touch the obstacle, its end is free to
move. More precisely, we have u0 (0, t) = 0 on the set {t : u(0, t) > 0}.
We suppose that the initial displacement u0 belongs to the Sobolev space H1 (0, L) and satisfies
the compatibility conditions, i.e. u0 (L) = 0 and u0 (0) ≥ 0 and the initial velocity v 0 belongs to
L2 (0, L).
We describe now the weak formulation of the problem. To this aim, we denote by K the
following convex set:
def
K = {u ∈ H2 : u(·, t) ∈ K for almost every t},
where H2 = {u ∈ L2 (0, T ; V ) : u˙ ∈ L2 (0, T ; L2 (0, L))} and K = {u ∈ V : u(0) ≥ 0} with
def
V = {u ∈ H1 (0, L) : u(L) = 0}.
def
def
Then the weak formulation associated to (1.1)–(1.3) is obtained by multiplying (1.1) by v −u,
def
v ∈ K and by integrating formally this result over QT = (0, L) × (0, T ):


find u ∈ K such that

 ∫ L
∫
∫

0
0
(1.4)
−
v (v(·, 0)−u ) dx −
u(
˙ v−
˙ u)
˙ dx dt +
u0 (v 0 −u0 ) dx dt ≥ 0

0
QT
QT



for all v ∈ K for which there exists ζ > 0 with v = u for t ≥ T − ζ.
24
1.2. Existence and uniqueness results by using the characteristic method
The weak formulation (1.4) is derived from [89] where the contact conditions are given in a
slightly different context. Existence and uniqueness results are obtained for a vibrating string
with concave obstacle in one-dimensional space in [40] and for a wave equation with unilateral
constraint at the boundary in a half space of RN in [44]. An existence result for a wave equation
in a C2 regular bounded domain constrained by an obstacle at the boundary in RN for N ≥ 2 is
proven in [45].
The paper is organized as follows. In Section 1.2, the problem (1.1)–(1.3) is reformulated as
a differential inclusion problem by using characteristic method, which is a crucial ingredient to
prove the uniqueness result. Then, the rest of this section is devoted to the proof of existence and
uniqueness results as well as to the energy balance. In Section 1.3, the equivalence between the
weak formulation associated to (1.1)–(1.3) and the differential inclusion obtained in Section 1.2 is
established. Then, in Section 1.4, a mass redistribution method is introduced and its convergence
is proved. This method is based on a redistribution of the body mass such that there is no inertia
at the contact node (see [84, 90]). Finally, some numerical examples are reported and analyzed
in Section 1.5. More precisely, the energy with and without the mass redistribution method are
compared as well as the approximated solution associated to the mass redistribution method and
an exact solution.
1.2
Existence and uniqueness results by using the characteristic
method
This section is devoted to the proof of existence and uniqueness results for the problem (1.1)–
(1.3). The first step consists in rewriting (1.1)–(1.3) as a differential inclusion problem by using
the characteristic method. To this aim, we introduce the following notations:
def
def
ξ = x + t and η = x − t.
Therefore the chain rule gives
∂2u
∂2u
∂2u
∂2u
+
=
+
2
∂x2
∂ξ 2
∂ξ∂η ∂η 2
which by using (1.1) implies that
∂2u
∂ξ∂η
and
∂2u
∂2u
∂2u
∂2u
+
=
−
2
,
∂t2
∂ξ 2
∂ξ∂η ∂η 2
vanishes. Thus we may conclude that
u(ξ, η) = p(ξ) + q(η),
where p and q are two differentiable functions such that
(1.5)
u(x, t) = p(x+t) + q(x−t).
In particular, taking t = 0 and using the initial data u0 and v 0 , we get
(1.6)
p(x) + q(x) = u0 (x) and p0 (x) − q 0 (x) = v 0 (x),
which by integration gives
(1.7)
u0 (ξ) 1
p(ξ) =
+
2
2
∫
ξ
u0 (η) 1
v (¯
x) d¯
x and q(η) =
−
2
2
∫
0
0
25
η
v 0 (¯
x) d¯
x
0
Chapter 1. Convergence of mass redistribution method for the one–dimensional
wave equation with a unilateral constraint at the boundary
for all ξ and η belonging to [0, L]. According to (1.5), the boundary conditions (1.3) can be
rewritten as follows:
(1.8)
0 ≤ p(t) + q(−t) ⊥ p0 (t) + q 0 (−t) ≤ 0 and p(L+t) + q(L−t) = 0
for all t belonging to [0, T ]. Thanks to the above identity, we may extend p(t) for all t ∈ [L, 2L],
i.e. we have
p(L+t) = −q(L−t)
for all t belonging to [0, L]. If we choose t0 = L + t, we get p(t0 ) = −q(2L−t0 ). We already have
the solution for q(t) with 0 ≤ t ≤ L and if L ≤ t0 ≤ 2L, we can obtain p(t0 ) by observing that
0 ≤ 2L − t0 ≤ L and by using q(2L−t), it follows that
∫
u0 (2L−t) 1 2L−t 0
(1.9)
p(t) = −
+
v (¯
x) d¯
x
2
2 0
for all t belonging to [L, 2L].
¯ defined by
Let us introduce the multivalued function JN : R → P(R)


if x < 0,

{0}
def
JN (x) = [0, +∞) if x = 0,


∅
if x > 0,
¯ is the set of all subsets of R.
¯ More precisely, JN (x) is the subdifferential of the
where P(R)
indicator function ∂I(−∞,0] (x) defined by
{
if x ∈ (−∞, 0],
def 0
I(−∞,0] (x) =
+∞ if x ∈
/ (−∞, 0].
Obviously, I(−∞,0] is a lower semi-continuous and convex function, for further details the reader
is referred to [91]. Then, the inequalities in (1.8) can be rewritten as follows
(1.10)
p0 (t) + q 0 (−t) ∈ −JN (−p(t)−q(−t))).
Note that at this stage, q(η), η ∈ [−L, 0], is the unique unknown of (1.10). We define now
def
f (t) = −p(t) − q(−t).
(1.11)
We insert (1.11) into (1.10) to get
f 0 (t) ∈ −JN (f (t)) − 2p0 (t).
Finally, we find the following Cauchy problem
(1.12a)
f 0 (t) ∈ −JN (f (t)) − 2p0 (t) a.e. t ∈ (0, L),
(1.12b)
f (0) = −u0 (0).
Observe that the Cauchy problem (1.12) formally is equivalent to (1.1)–(1.3). Note that the
existence and uniqueness results in half-space, with some appropriate conditions on u0 and v 0 ,
were established in [44]. The proof of Theorem 1 is rather classical. However for the reader
convenience, this proof is given in the Appendix A.
26
1.2. Existence and uniqueness results by using the characteristic method
Theorem 1 (Existence and uniqueness results). Assume that p belongs to W1,1 (0, L). Then the
Cauchy problem (1.12) admits a unique absolutely continuous solution.
˚ and ∂X be the interior and boundary of the
We introduce now some new notations: let X
set X, respectively, and let
I = {t ∈ [0, L] : f (t) = 0}
def
and
J = {t ∈ [0, L] : f (t) < 0}.
def
In the sequel, the notations for the constants introduced in the proofs are only valid in the proof.
The aim of the next lemma is to prove further regularity results for the solution f of Problem
(1.12).
Lemma 1 (Regularity result). Assume that p belongs to H1 (0, L). Then the solution f to Problem
(1.12) belongs to H1 (0, L).
Proof. Note that Theorem 1 implies that f is bounded in W1,1 (0, L). It follows that
f 0 (t) = 0 on ˚
I and f 0 (t) = −2p0 (t) on J.
Clearly, we have
∫
∫
0
|f (t)| dt =
J
∫
0
4|p (t)| dt < ∞ and
2
2
J
|f 0 (t)|2 dt = 0.
˚
I
Observe that if t is an accumulation point of ∂I, we may deduce that there exists a sequence
tn belonging to I such that tn → t so that f (tn ) = 0 and f (t) = 0. We recall the fundamental
theorem of calculus for Lebesgue integral states that if f is an absolutely continuous function on
[a, b], f is of bounded variation on [a, b]. Consequently, f 0 (t) exists for almost every t belonging
to [a, b]. For further details the reader is referred to [92, p. 160]. Hence f 0 vanishes for almost
all accumulation points of ∂I. It follows that f 0 (t) vanishes for almost every t belonging to ∂I.
Therefore, we deduce that f 0 is bounded in L2 (0, L), which implies that f belongs to H1 (0, L).
This concludes the proof.
It is convenient to define the following spaces:
def
H = L2 (0, L)
and
V = {u ∈ H1 (0, L) : u(L) = 0}
def
endowed with the norms k·kH and k·kV . Let (·, ·) and a(·, ·) the scalar products in H and V ,
respectively. This allows to define
H2 = {u ∈ L2 (0, T ; V ) : u˙ ∈ L2 (0, T ; H)}
def
endowed with the norm
kukH2 =
def
(∫
T(
) )1/2
ku(·, t)k2V +ku(·,
˙ t)k2H dt
,
0
and hh·, ·ii the duality corresponding product between H02 and H2 . We observe that H2 ,→
C0 ([0, T ]; H) (see [93]).
27
Chapter 1. Convergence of mass redistribution method for the one–dimensional
wave equation with a unilateral constraint at the boundary
Lemma 2. Assume that u belongs to H2 and u
¨ − u00 (defined in the sense of distributions on QT )
is square integrable. Then we have for all ε > 0, u ∈ C0 ([0, T ]; H1 (ε, L)) ∩ C0 ([0, L]; H1 (ε, T −ε)),
u˙ ∈ C0 ([0, T ]; L2 (ε, L)) and u0 ∈ C0 ([0, L]; L2 (ε, T −ε)).
Proof. The proof is obtained by using the same techniques detailed in [89]. Since it is quite a
routine to adapt this proof to our case, we let the verification to the reader.
The aim of the next lemma is to obtain some further regularity results for the solution u to (1.4).
Lemma 3. Let u be the solution to (1.4). Then for all ε > 0, u ∈ C0 ([0, L]; H1 (0, T −ε)) and
u0 ∈ C0 ([0, L]; L2 (0, T −ε)).
Proof. The proof of this result exploits the local energy identity inside QT , the reader can find
a detailed proof in the Appendix as well as in [89] where a succinct proof is given.
We deal now with the energy balance. More precisely, we prove below that the energy associated
to (1.12) given by
∫
)
1 L( 0
|u (t, x)|2 +|u(t,
˙ x)|2 dx
E(t) =
2 0
is constant with respect to time t.
Lemma 4 (Energy balance). Assume that p is bounded in W1,1 (0, L). Then the solution u to
Problem (1.12) is energy conserving.
Proof. We observe first that (1.5) gives
∫
)
1 L( 0
(p (x+t)+q 0 (x−t))2 +(p0 (x+t)−q 0 (x−t))2 dx
E(t) =
2 0
∫ L
∫ L
0
2
|q 0 (x−t)|2 dx.
=
|p (x+t)| dx +
(1.13)
0
0
We evaluate now the two integrals on the right hand side of (1.13). We note first that (1.7),
(1.9) and (1.11) lead to
{ (
1
0
p (x+t) =
)
00
0
2 u (x+t)+v (x+t) ,
( 00
)
1
0
2 u (2L−(x+t))−v (2L−(x+t)) ,
and
{ (
1
0
q (x−t) =
(
2
)
0
u0 (x−t)−v 0 (x−t) ,
f 0 (−(x−t))+p0 (−(x−t))
)
0 ≤ x + t ≤ L,
L ≤ x + t ≤ 2L,
0 ≤ x − t ≤ L,
, −L ≤ x − t ≤ 0.
On the one hand, we may deduce that
∫
(1.14)
0
L
∫
1 L 00
|u (x+t)+v (x+t)| dx +
|u (2L−x−t)−v 0 (2L−x−t)|2 dx
4 L−t
0
∫
∫
1 L 00
1 L 00
|u (x)+v 0 (x)|2 dx +
|u (x)−v 0 (x)|2 dx.
=
4 t
4 L−t
1
|p (x+t)| dx =
4
0
2
∫
L−t
00
0
2
28
1.3. The equivalence between the variational formulation and the differential
inclusion
On the other hand, by using the same kind of arguments as in the proof of Lemma 1, we may
obtain
{
|0+p0 (−η)|2
a.e. on I,
|q 0 (η)|2 = |f 0 (−η)+p0 (−η)|2 =
0
0
2
|−2p (−η)+p (−η)|
on J.
for all η belonging to [−L, 0]. It follows that |q 0 (η)|2 = |p0 (−η)|2 almost everywhere on [−L, 0].
Hence we have
∫
∫ L
∫ t
1 L 00
0
2
0
2
|u (x−t)+v 0 (x−t)|2 dx
|q (x−t)| dx =
|p (t−x)| dx +
4 t
0
0
∫
∫
1 t 00
1 L 00
(1.15)
=
|u (t−x)+v 0 (t−x)|2 dx +
|u (x−t)+v 0 (x−t)|2 dx
4 0
4 t
∫
∫
1 L−t 0 0
1 t 00
0
2
|u (x)+v (x)| dx +
|u (x)−v 0 (x)|2 dx.
=
4 0
4 0
Inserting (1.14) and (1.15) into (1.13), we get
∫
∫
1 L 00
1 L 00
E(t) =
|u (x)+v 0 (x)|2 dx +
|u (x)−v 0 (x)|2 dx,
4 0
4 0
and the conclusion is clear.
1.3
The equivalence between the variational formulation and the
differential inclusion
The present section is dedicated to prove the equivalence between the weak formulation (1.4) and
the differential inclusion (1.12). Consequently together with the results obtained in the previous
section, it is possible to deduce that (1.4) possesses a unique solution.
Lemma 5. Assume that u0 and v 0 belong to K = {u ∈ V : u(0, ·) ≥ 0} and H, respectively.
Then the unique solution to Problem (1.12) defines a weak solution to (1.4) for T = L.
def
Proof. The idea of the proof consists to split the domain QT into four regions according to
Figure 1.2 and to use the expression of the solution on each region to show that u and u˙ belong
to L2 (0, T ; V ) and to L2 (0, T ; H), respectively (see Fig. 1.2).
Let us go into the details. On the one hand, we observe that x + t ∈ [0, L] and x − t ∈ [−L, 0]
in the region I while x + t ∈ [L, 2L] and x − t ∈ [−L, 0] in the region II and we have
(1.16a)
u(x, t) = p(x+t) − p(t−x) − f (t−x),
(1.16b)
u0 (x, t) = p0 (x+t) + p0 (t−x) + f 0 (t−x),
(1.16c)
u(x,
˙
t) = p0 (x+t) − p0 (t−x) − f 0 (t−x).
Since u0 ∈ K, v 0 ∈ H and (1.7)–(1.9) hold, we may infer that p belongs to H1 (0, 2L) in the
regions I and II. Besides, Lemma 1 implies that f belongs to H1 (0, L). According to (1.16), we
conclude that
∫ L∫ L
(
)
(1.17)
|u(x, t)|2 +|u0 (x, t)|2 +|u(x,
˙
t)|2 dt dx < +∞.
0
x
29
Chapter 1. Convergence of mass redistribution method for the one–dimensional
wave equation with a unilateral constraint at the boundary
t
(L, L)
(0, L)
II
I
III
IV
x
(L, 0)
Figure 1.2: Four regions allowing to determine the value of u.
On the other hand, we note that x + t ∈ [L, 2L] and x − t ∈ [0, L] in the region III while
x + t ∈ [0, L] and x − t ∈ [0, L] in the region IV and we have
(1.18a)
u(x, t) = p(x+t) + q(x−t),
(1.18b)
u0 (x, t) = p0 (x+t) + q 0 (x−t),
(1.18c)
u(x,
˙
t) = p0 (x+t) − q 0 (x−t).
Still using the fact that u0 ∈ K, v 0 ∈ H and (1.7)–(1.9) hold, we may infer that p and q belong
to H1 (0, L) in the regions III and IV. Thanks to (1.18), we may infer that
∫ L∫ x
(
)
|u(x, t)|2 +|u0 (x, t)|2 +|u(x,
˙
t)|2 dt dx < +∞.
(1.19)
0
0
Therefore, it follows from (1.17) and (1.19) that u and u˙ are bounded in L2 (0, T ; V ) and
L2 (0, T ; H), respectively. We deduce from Lemma 2 that u(0, ·) and u0 (0, ·) belong to L2 (0, T )
and to L2 (ε, T −ε) for all ε > 0, respectively, and u(·, 0) and u(·,
˙ 0) belong to L2 (0, L) and to
L2 (ε, L) for all ε > 0, respectively. It remains to verify that (1.4) holds. To this aim, we observe
by using the notations introduced above that
∫ L
∫
0
−
v (v(x, 0)−u(x, 0)) dx −
u(x,
˙
t)(v−
˙ u)(x,
˙
t) dx dt
0
∫
QT
0
0
0
u (x, t)(v −u )(x, t) dx dt = −
+
QT
∫
(1.20)
∫
−
L
v 0 (v(x, 0)−u(x, 0)) dx
0
(p0 (x+t)−q 0 (x−t))(v(x,
˙ t)−(p0 (x+t)−q 0 (x−t))) dx dt
QT
∫
+
(p0 (x+t)+q 0 (x−t))(v 0 (x, t)−(p0 (x+t)+q 0 (x−t))) dx dt
QT
for all v belonging to K such that there exists ζ > 0 with v = u for t ≥ L − ζ. We evaluate each
integral on the right hand side of (1.20). Thanks to (1.6), we have
∫ L
∫ L
0
(1.21)
v (v(x, 0)−u(x, 0)) dx =
(p0 (x)−q 0 (x))(v(x, 0)−(p(x)+q(x))) dx.
0
0
30
1.3. The equivalence between the variational formulation and the differential
inclusion
The second integral on the right hand side of (1.20) is integrated by parts with respect to t, we
get
∫
(p0 (x+t)−q 0 (x−t))(v(x,
˙ t)−(p0 (x+t)−q 0 (x−t))) dx dt
QT
∫ L
=−
(p0 (x)−q 0 (x))(v(x, 0)−(p(x)+q(x))) dx
∫ 0
−
(p00 (x+t)+q 00 (x−t))(v(x, t)−(p(x+t)+q(x−t))) dx dt,
(1.22)
QT
while the third one is integrated by parts with respect to x, we find
∫
(p0 (x+t)+q 0 (x−t))(v 0 (x, t)−(p0 (x+t)+q 0 (x−t))) dx dt
QT
∫
T
=−
(p0 (t)+q 0 (−t))(v(0, t)−(p(t)+q(−t))) dt
0
∫
(
(p00 (x+t)+q 00 (x−t))(v(x, t)− p(x+t)+q(x−t))) dx dt.
−
(1.23)
QT
We substitute (1.21)–(1.23) into (1.20) and according to (1.8), we find
∫ T
−
(p0 (t)+q 0 (−t))(v(0, t)−(p(t)+q(−t))) dt ≥ 0,
0
which implies that (1.4) holds.
Lemma 6. The weak solution (1.4) for T = L defines a solution to Cauchy problem (1.12).
Proof. Let u be a solution to (1.4), it follows from Lemma 2 that u˙ has traces in L2 (ε, L) on
{t = 0} × (ε, L) and it comes that v 0 = u(·,
˙ 0) makes sense. We choose ψ = v − u such that ψ
belongs to the space of infinitely differentiable functions on QT with a compact support D(QT ).
Then it follows that
∫ T
∫ T
˙ dt +
−
(u,
˙ ψ)
a(u, ψ) dt = 0,
0
0
for all ψ belonging to D(QT ) where a(u, ψ) =
def
∫L
0
u0 ψ 0 dx. This gives
h¨
u, ψi − hu00 , ψi = 0
for all ψ belonging to D(QT ). Here the duality product between D0 (QT ) and D(QT ) is denoted
by h·, ·i. Therefore, we may deduce that u
¨ − u00 vanishes in the sense of distributions in QT .
Thus we have
∫
˙ 0 ψ 0 ) dx dt = 0
(1.24)
(−u˙ ψ+u
QT
for all ψ belonging to D(QT ). We introduce the following notations: µ+ = x + t, µ− = x − t
and ∆ denotes the region bounded by the lines µ+ = −µ− , µ+ = 2L − µ− , µ− = µ+ and
µ− = µ+ − 2T in the plane (µ− , µ+ ). Hence we have
def
(1.25)
u˙ =
∂u
∂u
−
,
∂µ+ ∂µ−
u0 =
∂u
∂u
+
,
∂µ+ ∂µ−
∂ψ
∂ψ
ψ˙ =
−
∂µ+ ∂µ−
31
and ψ 0 =
def
∂ψ
∂ψ
+
.
∂µ+ ∂µ−
Chapter 1. Convergence of mass redistribution method for the one–dimensional
wave equation with a unilateral constraint at the boundary
Carrying (1.25) into (1.24), we find
∫ (
∫ (
∂u
∂u )( ∂ψ
∂ψ )
∂u
∂u )( ∂ψ
∂ψ )
−
−
dµ− dµ+ +
+
+
dµ− dµ+ = 0
−
∂µ+ ∂µ−
∂µ+ ∂µ−
∆ ∂µ+ ∂µ−
∆ ∂µ+ ∂µ−
for all ψ belonging to D(∆). We observe that
∫
∫
∂u ∂ψ
∂u ∂ψ
(1.26)
dµ− dµ+ +
dµ− dµ+ = 0
∆ ∂µ+ ∂µ−
∆ ∂µ− ∂µ+
for all ψ belonging to D(∆). The first term in (1.26) is integrated by parts with respect to µ−
while the second one is integrated by parts with respect to µ+ to get
D
−
E D ∂2u
E
∂2u
,ψ −
, ψ = 0,
∂µ+ ∂µ−
∂µ− ∂µ+
for all ψ belonging to D(∆). Since we have
D
E D
E
∂2u
∂2ψ E D ∂2u
, ψ = u,
=
,ψ ,
∂µ+ ∂µ−
∂µ+ ∂µ−
∂µ− ∂µ+
it follows that
D
E
∂2u
,ψ = 0
∂µ+ ∂µ−
for all ψ belonging to D(∆). Then we conclude that
u = p(µ+ ) + q(µ− ).
∂2u
∂µ+ ∂µ−
vanishes which holds if and only if
We observe that Lemma 2 and Lemma 3 imply that u(·,
˙ 0) = v 0 and u0 (0, ·) belong to H
and to L2 (0, T −ε), respectively. According to Theorem 3 (given in Appendix B), the following
Green’s formulas make sense
∫ T
∫ T
∫ T∫ L
(
)
00
˙
(u,
˙ ψ)−a(u, ψ) dt +
u0 (L, t)ψ(L, t) dt
(¨
u−u )ψ dx dt = −
0
0
0
0
(1.27)
∫ T
(
) (
)
−
u0 (0, t)ψ(0, t) dt + u(x,
˙
T ), ψ(x, T ) − v 0 , ψ(x, 0)
0
for all ψ belonging to H2 . We insert (1.27) into (1.4) and we choose v = u + ψ, we obtain
∫
−
(1.28)
T
u0 (0, t)(v(0, t)−u(0, t)) dt ≥ 0
0
for all v belonging to K. Thanks to (1.5), we may deduce that (1.28) is equivalent to
∫
(1.29)
−
T
(p0 (t)+q 0 (−t))(v(0, t)−(p(t)+q(−t))) dt ≥ 0
0
for all v belonging to K. Since p(t) + q(−t) ≥ 0, it follows that
v(0, t) = p(t) + q(−t) + α(t)
32
1.4. Convergence of mass redistribution method
for all α(t) ≥ 0. Therefore we may infer from (1.29) that
∫
T
−
(p0 (t)+q 0 (−t))α(t) dt ≥ 0
0
for all α(t) ≥ 0, which implies that p0 (t) + q 0 (−t) ≤ 0 for almost every t ∈ (0, T ). Finally we
choose v(0, t) = 0 and v(0, t) = 2(p(t)+q(−t)) in (1.29), we get
∫
T
(p0 (t)+q 0 (−t))(p(t)+q(−t)) dt = 0,
0
which allows us to infer that (p0 (t)+q 0 (−t))(p(t) + q(−t)) vanishes for almost every t ∈ (0, T ).
This concludes the proof.
1.4
Convergence of mass redistribution method
The semi-discretized problem by using finite elements is not well-posed which emphasizes some
instabilities of time integration schemes (see [94, 95, 96, 97, 84]). In the literature many different
approaches were elaborated to overcome this difficulty. For instance the uniqueness for an impact
law of rigid bodies can be recovered by introducing a restitution coefficient (see [85]). However,
this approach is not totally satisfactory for deformable bodies. Indeed the presence of oscillations
due to displacement and to normal stress on the contact boundary induces some difficulties in
the construction of energy conserving schemes (see [78, 85, 80]). Another approach consists in
using the penalty method which introduces some oscillations as well but which can be reduced
with a damping technique (see [98]). One of the key points to avoid oscillations is to use the
mass redistribution method, the reader is referred to [84] and the references therein.
We first approximate (1.1)–(1.3) by using the Lagrange affine finite element method. To this
aim, we define h = Ln where n is an integer and
V h = {v h ∈ C0 ([0, L]) : v h |[ai ,ai+1 ] ∈ P1 , i = 0, . . . , n − 1, v h (L) = 0}.
def
def
Here, ai = ih, i = 0, 1, . . . , n, and P1 is the space of polynomials of degree less than or equal to 1.
A classical basis of V h is given by the sequence of shape functions ϕi ∈ V h for i = 0, 1, . . . , n − 1,
defined by
{
i|
1 − |x−a
if x ∈ [amax(i−1,0) , ai+1 ],
def
h
ϕi (x) =
0
otherwise.
Note that ϕi (aj ) = δij , j = 0, 1, . . . , n − 1, i.e. δij = 1 if i = j and δij vanishes otherwise (δ is
the Kronecker symbol). We approximate the solution u belonging to V to the weak formulation
(1.4) by
h
u (x, t) =
n−1
∑
j=0
33
uj (t)ϕj (x).
Chapter 1. Convergence of mass redistribution method for the one–dimensional
wave equation with a unilateral constraint at the boundary
Consequently, we have ui = uh (ai ), i = 0, 1, . . . , n − 1. The weak formulation (1.4) is approximated as follows


find uh : [0, T ] → V h and λ : [0, T ] → R such that for all v h ∈ V h


∫


 L h h

u
¨ v dx + a(uh , v h ) = −λv h (0) a.e. t ∈ [0, T ],
(Puh )
0



0
≤ uh (0, ·) ⊥ λ ≤ 0 a.e. t ∈ [0, T ],



uh (·, 0) = u0h and u˙ h (·, 0) = v 0h ,
where u0h and v 0h belong to V h such that
(
)
lim ku0h −u0 kV +kv 0h −v 0 kH = 0,
(1.30)
h→0
def
where λ is the Lagrange multiplier. We introduce now the following notations: U =
def
(u0 , . . . , un−1 )T and e0 = (1, 0, . . . , 0)T . The corresponding algebraic formulation is given by
(PU λ )


find U : [0, T ] → Rn and λ : [0, T ] → R such that



M U
¨ + SU = −λe0 a.e. t ∈ [0, T ],

0 ≤ u0 ⊥ λ ≤ 0 a.e. t ∈ [0, T ],




U (0) = U 0 and U˙ (0) = V 0 ,
where M and S denote the mass and stiffness matrices, respectively;

1 −1
0

−1
2 −1


∫ L
1  0 −1
2
def
def
ϕi ϕj dx
and
Sij = a(ϕi , ϕj ) = 
Mij =

.
.
.
h  ..
..
..
0

 ..
..
 .
.
0 ··· ···
···
..
.
..
.
..
.
−1
0
···
0
..
.
..
.





..

.
,

..
.
0


2 −1
−1
1
for all i, j ∈ [0, n − 1].
def ∫ L
We define now the modified mass matrix as follows: Mijmod = h ϕi ϕj dx. Clearly, we may
observe that M0i = Mi0 = 0 for all i = 0, . . . , n − 1. Therefore, the modified mass matrix reads
(
)
0 0
mod def
M
=
.
¯
0 M
¯ ij = Mi+1,j+1 for all i, j = 1, . . . , n − 2. We introduce now the following notations:
Note that M
def
def
def ∫
¯
U = (u1 , . . . , un−1 )T and S¯ij = Si+1,j+1 with C = Ω ϕ0i+1 ϕ00 dx, i = 0, . . . , n − 2. Observe that
C = (S10 , 0, . . . , 0)T . Thus by using the above notations, we have
(
S=
S00 C T
C
S¯
)
(
and
34
U=
u0
¯
U
)
.
1.4. Convergence of mass redistribution method
This leads to an algebraic formulation of the semi-discrete approximation with mass redistribution method given by

find U : [0, T ] → Rn and λ : [0, T ] → R such that





¯ ¨¯
¯¯


 M U + S U = −Cu0 a.e. t ∈ [0, T ],
¯ = −λ a.e. t ∈ [0, T ],
(Pmod
S00 u0 + C T U
Uλ )




0 ≤ u0 ⊥ λ ≤ 0 a.e. t ∈ [0, T ],




U (0) = U 0 and U˙ (0) = V 0 .
It follows that
u0 =
( −λ−C T U
¯)
S00
=
( −λ−S u )
10 1
.
S00
¯ )− otherwise we have
If S10 u1 ≥ 0 then the compatibility condition gives u0 = 0, so λ = (C T U
(
)+
u1
λ = 0. This implies that u0 = −SS10
, and then we may conclude that (Pmod
U λ ) is equivalent
00
to the following second order Lipschitz continuous ordinary differential equation:

¯ : [0, T ] → Rn−1 such that
find U




(
)+
¯U
¯¨ + S¯U
¯ = −C −S10 u1
(Pmod
)
M
a.e. t ∈ [0, T ],
¯
U

S00



U (0) = U 0 and U˙ (0) = V 0 .
mod
Lemma 7 (Existence and uniqueness results for (Pmod
U λ )). Problem (PU λ ) admits a unique
solution (U, λ) which is Lipschitz continuous.
¯ is not a singular matrix as well as the same techniques detailed
Proof. We use the fact that M
mod
in [99] to establish that (PU¯ ) possesses a unique Lipschitz continuous solution. On the other
−S10 u1 +
T¯ −
hand, we may deduce from (Pmod
U λ ) that u0 = ( S00 ) and λ = (C U ) . This allows us to
conclude that u0 and λ are also Lipschitz continuous and then the conclusion is clear.
We deal with the energy balance and we establish the energy conservation of the solution to
mod
problem (Pmod
U λ ) . More precisely, the discrete energy associated to problem (PU λ ) is given by
E(t) = 12 (U˙ T M mod U˙ +U T SU )(t).
(1.31)
Lemma 8. The solution (U, λ) to problem (Pmod
U λ ) is energy conserving.
Proof. We observe first that
¨ + U˙ T SU = −U˙ T λe0 .
U˙ T M mod U
Therefore, we integrate this expression over (0, t) to get
∫ t
∀t ∈ [0, T ] : E(t) − E(0) = −
u˙ 0 (s)λ(s) ds.
0
Let us define ω = {t ∈ [0, T ] : u0 (t) > 0}. On the one hand, the contact conditions imply that
λ = 0 on ω. On the other hand, the continuity of λ on [0, T ] gives that λ = 0 on ω
¯ where ω
¯ is
the closure to ω. Furthermore, u˙ 0 vanishes in the interior of the set [0, T ]\ω. Hence u˙ 0 λ = 0 on
[0, T ] and we conclude that E(t) = E(0) for all t ∈ [0, T ].
def
35
Chapter 1. Convergence of mass redistribution method for the one–dimensional
wave equation with a unilateral constraint at the boundary
We observe that (Pmod
U λ ) is equivalent to


find uh : [0, T ] → V h such that for all v h ∈ K h


∫ T (∫ L
)
h h
h
h h
h
(Pmod
)
u
¨ (v −u ) dx + a(u , v −u ) dt ≥ 0,
uh


0
h

 h
u (·, 0) = u0h and u˙ h (·, 0) = v 0h .
) to the solution of (1.4) by using
We establish below the convergence of the solution uh of (Pmod
uh
some ideas developed in [89].
Theorem 2. Assume that (1.30) holds. Then, the solution uh of (Pmod
) converges strongly in
uh
H2 to the unique solution of (1.4) as h tends to 0.
Proof. We observe that
∫ L
∫ h
∫
h
2
h
2
(1.32)
|u˙ (x, t)| dx =
|u˙ (x, t)| dx +
0
0
L
|u˙ h (x, t)|2 dx.
h
(
)+
10
We evaluate now the right hand side of (1.32). To this aim, we note that u0 (t) = − SS00
u1 (t)
implies that
|u˙ 0 (t)| ≤ |u˙ 1 (t)|,
(1.33)
10
since − SS00
= 1. Therefore, by using (1.33) and Cauchy Schwarz’s inequality, we may deduce
that there exists C0 > 0 such that
∫ h
∫ h
h
2
|u˙ 1 (t)ϕ1 (x)+u˙ 0 (t)ϕ0 (x)|2 dx ≤ C0 h|u˙ 1 (t)|2 .
|u˙ (x, t)| dx =
0
0
Furthermore, the energy conservation of Lemma 8 implies that there exists C1 > 0 such that
∫ L
|u˙ h (x, t)|2 dx ≤ C1 .
h
Consequently, we deduce that
∫ (i+1)h
∫
h
2
|u˙ (x, t)| dx =
ih
(i+1)h
|u˙ i+1 (t)ϕi+1 (x)+u˙ i (t)ϕi (x)|2 dx ≤ C1
ih
for i = 1, . . . , n − 1. We conclude that for sufficiently small h, we get
|u˙ i+1 (t)|2 + |u˙ i (t)|2 ≤
6C1
h
1
for i = 1, . . . , n − 1. Therefore |u˙ 1 (t)|2 ≤ 6C
˙ h (·, t)kL2 (0,h) is bounded
h which implies that ku
independently of h and then ku˙ h (·, t)kH is also bounded independently of h. By using Lemma
8, we can prove that kuh kV is bounded. It follows that there exists C > 0 independent of t such
that
sup (kuh (·, t)kV +ku˙ h (·, t)kH ) ≤ C.
t∈[0,T ]
36
1.4. Convergence of mass redistribution method
Then, it is possible to extract a subsequence, still denoted by uh , such that
uh * u
(1.34a)
h
(1.34b)
u˙ * u˙
L∞ (0, T ; V )
in
weak
∗,
L (0, T ; H) weak
∗.
∞
in
Let us define
H∞ = {u ∈ L∞ (0, T ; V ) : u˙ ∈ L∞ (0, T ; H)}
def
endowed with the following norm
(
)
def
kukH∞ = ess sup ku(·, t)kV + ku(·,
˙ t)kH .
t∈[0,T ]
We may infer from (1.34) that
uh * u
in
H∞
weak
∗.
1
Since for all α < 12 the following injections H∞ ,→ C0, 2 (QT ) ,→,→ C0,α (QT ) hold (see [89]),
where ,→ is continuous embedding and ,→,→ is compact embedding, we get
uh → u
in
C0,α (QT )
for all α < 21 . We observe that both uh (t) and u(t) belong to K.
In order to prove that the limit u satisfies (1.4), it is necessary to choose convenient test
functions. We approximate the elements of K before projecting them onto V h . Indeed, the
L2 projection does not conserve the constraint at x = 0, and therefore, the elements of K need
another approximation in order to satisfy the constraint strictly. To this aim, let v be an element
of K which is equal to u for t ≥ T − ε and let

∫ t+β

u(x, t) + 1
(v(x, s)−u(x, s)) ds + k(β)(L−x)g(t) if t ≤ T − β,
def
β t
v β (x, t) =


u(x, t)
if t ≥ T − β,
for all β ≤ 4ε . Here g(t) is a smooth, positive function (see Figure 1.3) and satisfying
g(t)
1
t
0
T − 2ǫ
T − 4ǫ T
Figure 1.3: Smooth and positive function g(t).
{
def
g(t) =
1
if
t ∈ [0, T − 2ε ],
0
if
t ∈ [T − 4ε , T ].
37
Chapter 1. Convergence of mass redistribution method for the one–dimensional
wave equation with a unilateral constraint at the boundary
1
The next step consists to choose adequately k(β). Since u belongs to C0, 2 (QT ), we may deduce
that there exists C > 0, such that
∫ t+β
∫ t+β
−1
−1
u(0, s) ds ≤ β
|u(0, t)−u(0, s)| ds
u(0, t) − β
t
t
∫ β
√
√
2
−1
s ds = CkukH∞ β.
≤ CkukH∞ β
3
0
Furthermore, we have the following inequality
∫ t+β
√
2
β
−1
v (0, t) ≥ β
v(0, s) ds − CkukH∞ β + k(β)Lg(t)
3
t
√
5
for all t ≤ T − 2ε . Choosing k(β) = 3L
CkukH∞ β, we get
√
v β (0, t) ≥ CkukH∞ β
for all t ≤ T − 2ε . On the other hand, we have
v β (x, t) = u(x, t) + k(β)(L−x)g(t)
for all t belonging to [T − 2ε , T −β]. Hence v β belongs to K and in the other hand v β belongs to
L∞ (0, T ; V ) because
1
kv β (·, t)−u(·, t)kV ≤ k(β)Lg(t) + √ kv−ukH2 .
β
We denote by Qh the orthogonal projection onto V h with respect to the scalar product in H
such that kQh z−zkV → 0 when h → 0 for all z ∈ V (see [100]). The Sobolev injections imply
that there exists a sequence αh converging to zero as h tends to zero such that
kQh z−zkC0 ≤ αh kzkV ,
for all z belonging to V with limh→0 αh = 0. The test function is defined as follows:
v h (·, t) = uh (·, t) + Qh (v β −u)(·, t)
def
(1.35)
for all t belonging to [0, T ]. By using a continuity argument, v h (0, ·) ≥ 0 for h small enough.
) and using the integration by parts, we find
Carrying (1.35) into (Pmod
uh
∫ L
∫ h
−
u˙ h (·, 0)Qh (v β −u)(·, 0) dx +
u˙ h (·, 0)Qh (v β −u)(·, 0) dx
0
0
∫ T∫ L
∫ T∫ h
h
h β
−
u˙ (·, t)Q (v˙ −u)(·,
˙
t) dx dt +
u˙ h (·, t)Qh (v˙ β −u)(·,
˙
t) dx dt
0
0
0
0
∫ T
+
a(uh (·, t), Qh (v β −u)(·, t) dt ≥ 0.
0
Since (v β −u)(·, t) is bounded in H∞ , the above integration makes sense. Thus we may pass to
the limit when h tends to zero. Since we have
Qh (v˙ β −u)
˙ → (v˙ β −u)
˙
in L2 (0, T ; H)
and
38
Qh (v β −u) → (v β −u) in L2 (0, T ; V ).
1.5. Numerical examples
Then, we conclude that
∫
−
∫
L
L
∫
v˙ 0 (v β −u)(·, 0) dx −
0
∫
T
u(·,
˙ t)(v˙ β −u)(·,
˙
t) dt dx +
0
0
T
a(u(·, t), (v β −u)(·, t)) dt ≥ 0.
0
We pass now to the limit with respect to β so we obtain variational formulation (1.4).
On the one hand, we observe that Lemma 4 leads to
∫ L
∫ L
0
2
0
2
(1.36)
(|u(·,
˙ t)| +|u (·, t)| ) dx =
(|v 0 |2 +|u0 |2 ) dx.
0
0
On the other hand, Lemma 8 implies that
∫ L
∫
h
2
h
h
|u˙ (·, t)| dx + a(u (·, t), u (·, t)) =
h
L
|u˙ h (·, 0)|2 dx + a(uh (·, 0), uh (·, 0)),
h
which by using (1.30) and (1.34) gives
∫
(1.37)
lim
L
h
h→0 h
h0
2
∫
L
2
(|u˙ (·, t)| +|u (·, t)| ) dx =
0
(|v 0 |2 +|u0 |2 ) dx.
h
Therefore from (1.36) and (1.37)
∫
lim
L
h→0 h
∫
0
(|u˙ h (·, t)|2 +|uh (·, t)|2 ) dx =
L
(|u(·,
˙ t)|2 +|u0 (·, t)|2 ) dx.
0
Since uh converges weakly to u in H∞ and kuh kH∞ converges to kukH∞ and since H∞ ,→ H2
then we conclude that uh converges strongly to u in H2 .
1.5
Numerical examples
We perform a finite element discretization in space and we use a classical Newmark time stepping
method. This leads to consider the following problem:

(
)
n+1 = U n + ∆tU
˙ n + 1 −β ∆t2 U
¨ n + β∆t2 U
¨ n+1 ,

U

2



˙ n+1 = U˙ n + (1−γ)∆tU
¨ n + γ∆tU
¨ n+1 ,


U
¨ n+1 + SU n+1 = −λn+1 e0 ,
(1.38)
MU




0 ≤ un+1
⊥ λn+1 ≤ 0,

0


U (0) = U 0 and U˙ (0) = V 0 ,
where β ∈]0, 1/2[ and γ ∈]1/2, 1[ are the classical parameters of the Newmark scheme. Note that
if β = 0.25 and γ = 0.5, the scheme (1.38) is the so-called Crank-Nicholson scheme which is an
implicit, unconditionally stable and second-order accurate scheme for elastodynamic problems
without contact conditions and moreover it is energy conserving (see [84]). On the other hand, it
is well known that the space-semi discretization of contact problems in elastodynamics presents
some numerical instabilities (see [59]) which can be avoided by using a modified mass method
(see [84] and the references therein). We make below some comparisons between two different
39
Chapter 1. Convergence of mass redistribution method for the one–dimensional
wave equation with a unilateral constraint at the boundary
approaches; the one using a standard mass matrix and the one using a modified mass matrix.
The parameters used in the numerical simulations are the space step ∆x = 0.1, the time step
∆t = 0.01, the initial displacement u0 (x) = 0.5x − 0.5, the initial velocity v 0 (x) = 0 and the
Dirichlet value u(L, t) = 0.45 with L = 1 and T = 5. The numerical experiments are performed
by employing the finite element library Getfem++ (see [87]). In particular, the generalized
Newton algorithm has been used to compute the unique solution of (1.38) (see [101, 102]). The
numerical results show that when the constraint is active, small oscillations occur in the case
where M is a standard mass matrix (see Figures 1.4 and 1.5 (left)) while these oscillations do not
exist in the case where M is a modified matrix (see Figures 1.4 and 1.5 (right)). Furthermore,
we can observe in Figure 1.6 (left), the energy is increasing with the standard mass matrix while
with modified mass matrix (right), it is almost conservative.
Figure 1.4: Numerical experiments with standard mass matrix (left) and with modified mass
matrix (right).
Figure 1.5: Numerical experiments with standard mass matrix (left) and with modified mass
matrix (right) in the contact point x = 0.
40
1.5. Numerical examples
Figure 1.6: Energy evolution for standard mass matrix (left) and for modified mass matrix
(right).
We present now some numerical results obtained for an undeformed elastic bar which is dropped
with some initial velocity on a rigid obstacle and we compare the exact solution to the approximated one obtained by using the mass redistribution method. More precisely, we assume that
this bar falls from a height u0 , with an initial velocity −v 0 and under the gravity g ≥ 0. Furthermore both ends of the bar are free to move, as long as the bar does not hit a rigid obstacle.
The length and the Young modulus of the bar are denoted by L and E, respectively. Let u(x, t)
be the displacement at time t of the material point of spatial coordinate x ∈ [0, L] and the
contact pressure equal to the normal stress −Eu0 (0, ·). Then the mathematical problem can be
formulated as follows:
(1.39)
u
¨(x, t) − Eu00 (x, t) = −(g+Eu0 (0, t)),
(x, t) ∈ (0, L) × (0, T ),
with Cauchy initial data
(1.40)
u(·, 0) = u0
and
u(·,
˙ 0) = −v 0 ,
and Signorini and Neumann boundary conditions at x = 0 and x = L, respectively, for t > 0
(1.41)
0 ≤ u(0, ·) ⊥ Eu0 (0, ·) ≤ 0
and
u0 (L, ·) = 0.
The existence and uniqueness results for (1.39)–(1.41) is obtained by rewriting this problem as
a differential inclusion problem and then by using the same techniques detailed in the proof of
Theorem 1. Since it is quite a routine to adapt this proof to the case considered here, we leave
the verification to the reader. In order to calculate the analytical solution to problem (1.39)–
(1.41), we distinguish three phases, namely before the contact, during the contact and after the
contact. To this aim, we choose v 0 = 0 and g >
impacts.
√0 so that the bar can make several
√
2u0
0
The bar reaches the rigid obstacle at time t1 =
2u g. After
g with the velocity equal to
the impact, the bar stays in contact with the rigid body; as soon as a shock wave travels from
bottom of the bar to the top and vice versa then the bar takes off. The wave takes a time √LE
41
Chapter 1. Convergence of mass redistribution method for the one–dimensional
wave equation with a unilateral constraint at the boundary
to travel along the bar. The impacts occur at time t4k+1 = 3 √LE + 16k √LE , t4k+2 = t4k+1 + 2 √LE ,
t4k+3 = t4k+1 + 8 √LE , t4k+4 = t4k+1 + 10 √LE . We introduce also the following notations:
√
∞
( x
√
2u0
L L ) ∑ 2g
min √ , √ −t− √ +
(cos( Eνn t)−1) sin(νn x),
3
g
ELνn
E
E
E
n=1
√
∞
√
2u0 )2 2gL2 ∑ 4g
1 (
cos( Eλn t) cos(λn x),
h2 (x, t) = u0 − g t−
−
+
2
2
g
3E
Eλn
h1 (x, t) = −
n=1
π
with νn = (n− 21 ) L
and λn = n Lπ . Then, the explicit solution reads as
(1.42)

1 2
0

u − 2 gt




h (x, t−t4k+1 )

 1
u(x, t) = h2 (x, t−t4k+2 )



h1 (x, t4k+4 − t)


√ )2
(


u0 − 1 g t−t
2u0
4k+4 −
2
g
if
t ≤ t1 ,
if
t4k+1 < t ≤ t4k+2 ,
if
t4k+2 < t ≤ t4k+3 ,
if
t4k+3 < t ≤ t4k+4 ,
if
t4k+4 < t ≤ t4(k+1)+1 .
Here, some details are omitted, the reader is referred to [103, 9] for a detailed explanation. We
choose L = 10, T = 6.5, E = 900, g = 10, the initial data u0 (x) = 5, v 0 (x) = 0 and Neumann
value u0 (L, t) = 0. The Newmark time stepping method with β = 0.25 and γ = 0.5 is used to
evaluate the approximated solution. Let us emphasize that if the space step ∆x and time step ∆t
tend to 0, the approximated solution obtained by using the mass redistribution method (Pmod
Uλ )
converges to the solution of (1.39)–(1.41) explicitly given by (1.42) (see Figure 1.7 (left)). On
the other hand, we can write at least formally an energy relation for (1.39): we multiply this
equation by u,
˙ we integrate by parts over Qτ , τ ∈ [0, T ], we get
1
2
∫
L
1
|u(·,
˙ τ )| dx +
2
∫
2
0
L
∫
√ 0
2
| Eu (·, τ )| dx = −
0
g u˙ dx dt
Qτ
for all τ belonging to [0, T ]. Observe that the energy tends to be conserved when the space step
∆x and the time step ∆t tend to 0 and the energy decreases otherwise (see Figure 1.7 (right)).
42
1.5. Numerical examples
Figure 1.7: Numerical convergence of the solution associated to problem with mass redistribution
method to the exact solution in the contact point x = 0 (left). Numerical convergence of the
energy evolution associated to mass redistribution method to the exact energy (right).
Appendix A
The aim of this section is to give the proofs of Theorem 1 and Lemma 3. Furthermore, a regularity
result is also presented. Notice that Theorem 1 is a straightforward application of [104, p. 59
Cor. 5.4].
def
Proof of Theorem 1. We verify the assumptions of [104, p. 59 Cor. 5.4]. We define F (t, f (t)) =
−JN (f (t)) − 2p0 (t). Hence we choose f (t) = y which gives that
F (t, y) = −JN (y) − 2p0 (t).
¯ has closed convex values and F is measurable
The multivalued map F : [0, L] × (−∞, 0] → P(R)
with respect to its second variable. We prove now that F is upper semi–continuous with respect
to its second variable which is equivalent to establish that JN is upper semi–continuous. Note
¯ J −1 (A) = [0, +∞) or J −1 (A) = {0} or J −1 (A) = ∅ which are closed sets.
that if A ⊂ R,
N
N
N
According to Definition [104, p.4, Def. 1.1], JN is upper semi–continuous.
We verify that there is a function r(t, y) = c(t)(1+|y|) with c ∈ L1 (0, L) such that
¯1 (0) ∩ T(−∞,0] (y) 6= ∅
F (t, y) ∩ r(t, y)B
on
[0, L] × (−∞, 0],
where B1 (0) is the ball of radius 1 at the origin and T(−∞,0] (y) is the tangent cone on y, the
reader is referred to [104] for further details. Indeed, we distinguish two cases, on the one hand,
if y belongs to the interior of (−∞, 0], T(−∞,0] (y) = R and F (t, y) = −2p0 (t), we choose c(t) =
1 + |2p0 (t)| and on the other hand, if y = 0, T(−∞,0] (0) = (−∞, 0] and F (t, y) = (−∞, −2p0 (t)],
we choose c(t) = 1 + |2p0 (t)|. Therefore the existence of solution to (1.12) follows.
The uniqueness result comes from the monotonicity of JN , namely JN is the subdifferential
of a convex, lower semi–continuous and proper function, the reader is referred to [91] for further
details. 43
Chapter 1. Convergence of mass redistribution method for the one–dimensional
wave equation with a unilateral constraint at the boundary
Proof of Lemma 3. We note that
(u02 +u˙ 2 )0 − 2
(1.43)
∂ 0
(u u)
˙ = 0 in the sense of distributions.
∂t
Hence we integrate (1.43) over [x0 , x1 ] × [t0 , t1 ], with 0 < x0 < x1 < L and 0 < t0 < t1 < T , to
get
∫ t1
∫ x1
02
2
02
2
(1.44)
((u +u˙ )(x1 , t)−(u +u˙ )(x0 , t)) dt =
((2u0 u)(x,
˙
t1 )−(2u0 u)(x,
˙
t0 )) dx.
t0
x0
According to Lemma 2, the right hand side of (1.44) is bounded independently of x0 , x1 , t0 , t1 as
x0 , L], we may
long as 0 < x
¯0 6 x0 < x1 < L. We integrate now (1.44) with respect to x0 over [¯
deduce that there exists C > 0 independent of x1 , x
¯0 such that
∫ t1
∫ L ∫ t1
(L−¯
x0 )
(u02 +u˙ 2 )(x1 , t) dt ≤
(u02 +u˙ 2 )(x0 , t) dt dx0 + C(L−¯
x0 ),
t0
x
¯0
∫ t1
t0
which implies that x 7→ t0 (u02 +u˙ 2 )(x1 , t) dt is bounded on [¯
x0 , L] independently of t0 and t1 , it
∫ T 02 2
follows that x 7→ 0 (u +u˙ )(x, t) dt is bounded on [¯
x0 , L]. Let v be the solution of the following
problem

00


v¨ − v = 0 on (0, x0 ) × (0, T ),
(1.45)
v(x0 , t) = u(x0 , t) and v 0 (x0 , t) = u0 (x0 , t) for
all t ∈ [0, T ],


v(x, 0) = u0 (x) and v(x, T ) = u(x, T ) for
all x ∈ [0, x0 ].
These conditions are illustrated in Figure 1.8. Since u0 (x0 , ·) and u(x
˙ 0 , ·) belong to L2 (0, T ),
v(x0 , t) = u(x0 , t)
x0
v ′ (x0 , t) = u′ (x0 , t)
v(x, 0) = u0 (x)
v(x, T ) = u(x, T )
v(x,
˙ 0) = v 0 (x)
0
T
Figure 1.8: Initial and boundary conditions for v on the rectangle (0, x0 ) × (0, T ).
we may infer that there exists a unique solution to (1.45). More precisely, w = u − v satisfies
(1.45) with homogeneous boundary initial conditions and the existence and uniqueness theorem
˙ 0) = v 0 (x).
in [105] holds. Furthermore v = u on (0, x0 ) × (0, T ) and in particular we have v(x,
We solve (1.45) by employing a classical characteristic method. To this aim, it is convenient to
introduce the following notations:
β1 = −x + t
def
and
2
β2 = −x − t.
def
v
vanishes which implies that v(x, t) = f (β1 ) + g(β2 ). Notice that the
We may deduce that ∂β∂1 ∂β
2
general solution for all of points in the rectangle (0, x0 ) × (0, T ) does not exist. Then we split the
44
1.5. Numerical examples
rectangle into three regions by using characteristic lines as it is shown on Figure 1.9. We look
for the solution in each region. More precisely, in region I and according to the initial condition
of problem (1.45) in x = x0 for x ≤ x0 , we get
∫
1
1 t+(x0 −x) 0
(1.46)
v(x, t) = (u(x0 , t+(x0 −x))+u(x0 , t−(x0 −x))) −
u (x0 , ζ) dζ.
2
2 t−(x0 −x)
Observe that Figure 1.9 gives a better interpretation of this phenomenon. Indeed, the interval
used will be the intersection of the line x = x0 with the forward wave cone at (x, t) which is the
region between the two straight lines having a slope of ±1 but directed upwards from an origin
(x, t). The forward wave cone at the point (x, t) will enclose all those points (x0 , ζ) which motion
will be influenced by what occurred at the point x at the time t. Concerning the region II, we
v(x0 , t) = u(x0 , t)
x0
v ′ (x0 , t) = u′ (x0 , t)
b
b
I
v(x, 0) = u0 (x)
II
v(x, T ) = u(x, T )
b
(x, t)
v(x,
˙ 0) = v 0 (x)
III
0
T
Figure 1.9: On characteristics in the region I.
use the characteristics illustrated by Figure 1.10. Let A = (xA , tA ) be a point in the region II.
v(x0 , t) = u(x0 , t)
x0
v ′ (x0 , t) = u′ (x0 , t)
b
I
v(x, 0) = u0 (x)
b
II
v(x,
˙ 0) = v 0 (x)
v(x, T ) = u(x, T )
III
b
A
0
T
Figure 1.10: Characteristics in the region II.
It follows that
(1.47a)
v 0 (xA , tA ) − v(x
˙ A , tA ) = u00 (xA +tA ) − v 0 (xA +tA ),
(1.47b)
v 0 (xA , tA ) + v(x
˙ A , tA ) = u0 (x0 , tA +(x0 −xA )) + u(x
˙ 0 , tA +(x0 −xA )).
Therefore (1.47) leads to
(1.48)
(
1
0
v 0 (xA , tA ) = (u0 (xA +tA )−v 0 (xA +tA )+u0 (x0 , tA +(x0 −xA ))+u˙ x0 , tA +(x0 −xA ))),
2
and
1
0
v(x
˙ A , tA ) = (−u0 (xA +tA )+v 0 (xA +tA ) + u0 (x0 , tA +(x0 −xA ))+u(x
˙ 0 , tA +(x0 −xA ))).
2
45
Chapter 1. Convergence of mass redistribution method for the one–dimensional
wave equation with a unilateral constraint at the boundary
We have the solution for all the points located in the regions I and II. Concerning the solution
in region III, we need some further regularity result to conclude. We obtained some regularity
results for u(x0 , t), u0 (x0 , t) and u(x
˙ 0 , t) in Lemma 2. Besides by using (1.48) and (1.46), it is
0
possible to deduce that v belongs to C0 ([0, x0 ]; L2 (0, T −x0 )) which implies that v belongs to
C0 ([0, x0 ]; H1 (0, T −x0 )). As x0 is arbitrary small, the conclusion is clear. Appendix B
We establish below a Green’s formula that is crucial in the proof of Lemma 6. To this aim, let us
introduce a linear topological space D and Hilbert spaces V, H, Z and S with their topological
duals denoted by D0 , V 0 , H0 , Z 0 and S 0 .
Suppose that D is contained in V and it is dense in H. Here H is identified with its dual,
namely we have H = H0 . Furthermore, Suppose that V is contained in H with finer topology
and we denote by V0 , the closure of D in V such that
D ,→ V0 ,→ H = H0 ,→ V00
with dense embedding. More precisely, D is an abstraction of the usual space D(QT ) of test
function. In our case the spaces V, S and Z denote the admissible displacement, stresses and
boundary values containing traces of element of V, respectively. We also introduce a linear
operator A ∈ L(V, S) and its restriction to V0 denoted by A0 ∈ L(V0 , S) such that
∀v ∈ V0 , Av = A0 v.
Let A∗ ∈ L(S 0 , V 0 ) be the adjoint of the operator A, defined by
hτ, AviS 0 ×S = hA∗ τ, viV 0 ×V .
Then
∀τ ∈ S 0 , ∀v ∈ V0 :
hτ, AviS 0 ×S = hA∗0 τ, viV 0 ×V .
Let us define S 0 (A∗0 ) = {τ ∈ S 0 : A∗0 τ ∈ H}, then we have the following trace theorem.
def
Theorem 3. Suppose that V0 is the kernel of a surjective map γ ∈ L(V, Z) from V onto Z. Then
there exists a unique linear operator π ∈ L(S 0 (A∗0 ), Z 0 ) such that the following Green’s formula
holds:
∀τ ∈ S 0 (A∗0 ), ∀v ∈ V : hτ, AviS 0 ×S − hA∗0 τ, viV 0 ×V = hπτ, γviZ 0 ×Z .
The detailed proof of (3) is given in [106]. In particular, we are interested in this work to the case
def
def
def
where S = S 0 = {(u1 , u2 ) ∈ L2 (QT )}, D = D(QT ), V = H1 (QT ), V0 = H10 (QT ), V00 = H−1 (QT ),
def
∂
∂
H = L2 (QT ), S 0 (A∗0 ) = {(u1 , u2 ) ∈ S : ∂t
u1 − ∂x
u2 ∈ L2 (QT ) in the sense of distributions},
def
∂
∂
A : u 7→ ( ∂t
u, − ∂x
u), Z = H1/2 (∂QT ) and Z 0 = H−1/2 (∂QT ), where (∂QT ) is the boundary of
QT and the trace operator γ : V → Z. Then there exists a unique π ∈ L(S 0 (A∗0 ), Z 0 ) such that
∫ (
∫ (
∂ )
∂
∂ )
∂
τ1 −
τ2 v dt dx = hπτ, γviZ 0 ×Z ,
τ1 v + τ2 v dt dx −
∂t
∂x
∂x
QT ∂t
QT
46
1.5. Numerical examples
for all v ∈ V and (τ1 , τ2 ) ∈ S 0 (A∗0 ). Then by density argument, we have
∫
hπτ, γviZ 0 ×Z =
(τ.n)v ds,
∂QT
where n is the outward unit normal to QT , when τ and v are regular enough.
Appendix C
Multivalued differential equations
In this Section, we recall some basic results for multivalued differential equations, the reader
refer to [107, 104] for further details. A multivalued map is a map which values are subsets of a
vector space X. Often these are particular subsets, for example closed convex subsets. These are
seen as extensions of ordinary functions on X, and elementary operations on multivalued map
are deduced from the elementary operations on X.
Definition 1. Let Xand Y be two Banach spaces. Then we denote P(X) the set of all subsets
of X. Given another set D 6= ∅ a subset of Y , a map F : D → P(X)\{∅} is called a multivalued
map.
A single–valued map means F (ω) = {f (ω)} on D. A function f : D → X with f (ω) ∈ F (ω) is
called a selection of F .
We describe now the main elementary operations on the multivalued maps. Firstly, for two
subsets A, B of X, we define
A + λB = {a + λb : a ∈ A, b ∈ B}
and
x + B = {x} + B.
So, for example, Br (x0 ) = x0 + rB1 (0) and
A + Br (0) = {x ∈ X : ρ(x, A) ≤ r}
ρ(x, A) = inf |x − a|.
with
a∈A
The addition between the multivalued maps is defined as follows:
(F + λG)(x) = F (x) + λG(x) = {f + λg : f ∈ F (x), g ∈ G(x)}.
Hausdorff distance between two sets A and B is defined by
{
}
dH (A, B) = max sup ρ(x, B), sup ρ(x, A) .
x∈A
x∈B
For A ⊂ X we denote by
kAk = sup kykX .
y∈A
The convex hull of A ⊂ X is the smallest convex set containing A, i.e.
convA =
m
{∑
i=1
λi xi : m ∈ N, λi ∈ [0, 1],
m
∑
}
λi = 1, xi ∈ A for i = 1, . . . , m .
i=1
The property of upper semicontinuity property of multivalued maps plays an important role to
obtain some existence results for differential inclusions.
47
Chapter 1. Convergence of mass redistribution method for the one–dimensional
wave equation with a unilateral constraint at the boundary
Definition 2. Let X, Y be Banach spaces and ∅ 6= D ⊂ Y , a map F : D → P(X)\{∅} is said
to be upper semi–continuous (usc), if F −1 (A) is closed in D whenever A ⊂ X is closed.
The condition of being usc may become more transparent in terms of sequences since it means:
if (ωn ) ⊂ D, A ⊂ X is closed, ωn → ω0 ∈ D and F (ωn ) ∩ A 6= ∅ for all n ≥ 1 then also
F (ω0 ) ∩ A 6= ∅. So, intuitively, F (ω0 ) must be at least as large as the limit of the F (ωn ).
A prototype of an usc multivalued map, defined by

for
t<0

 {0}
F (t) = {0, 1} for
t=0


{1}
for
t>0
Later, it will often be essential to have convex values, here, this can simply be achieved by letting
F (0) = [0, 1], i.e. by filling in the gap at the point of discontinuity. This situation can easily be
generalized, given f : D → X, by
∩
F (x) =
convf (Bδ (x) ∩ D)
for
x ∈ D.
δ>0
Obviously, f is a selection of F , F has closed convex values, and F (ω0 ) = {f (ω0 )} if f is
continuous at ω0 . However, if f is locally compact (i.e. for all ω0 ∈ D there exists r = r(ω0 ) > 0
such that f (Br (ω0 ) ∩ D) is compact) then F (x) is usc (see [104]).
Definition 3. [104, p.35 Sec 4.3] Let X be a real Hilbert space with inner product (·, ·), a map
F : D ⊂ X → P(X)\{∅} is said to be monotone if
(Fx − Fy , x − y) ≥ 0 on D × D,
which is understood as (u − v, x − y) ≥ 0 for all u ∈ Fx and all v ∈ Fy . In case −F is monotone
we say that F is dissipative.
˚ 6= ∅, and
Proposition 1. [104, p. 36 Prop. 4.2] Let X be a real Hilbert space, D ⊂ X with D
F : D → P(X)\{∅} be monotone. Then
˚
(a) F is locally bounded on D.
˚ dense in D,
˚ such that F | ˚ is single–valued.
(b) There exists a Gδ –set D0 ⊂ D,
D
Note that Gδ –set is the name for intersections of countably many open sets.
Definition 4. Let X be a Hilbert space with inner product (·, ·), F : D ⊂ X → P(X)\{∅} is
said semi–Lipschitz if there is k ∈ R such that
(F (x) − F (y), x − y) ≤ k k x − y k2X on D × D.
Given X = Rn , J = [0, T ] ⊂ R, be closed D ⊂ X, a multivalued map F : J × D → P(X)\{∅}
and x0 ∈ D, we are looking for absolutely continuous solution of
{
u(t)
˙
∈ F (t, u(t)) a.e. on J,
(1.49)
u(0) = u0 ∈ D.
48
1.5. Numerical examples
i.e. u : J → D such that
∫
u˙ ∈ L1 (J, Rn ) and u(t) = u0 +
t
u0 (s) ds on J
and u˙ ∈ F (t, u(t)) a.e. on J.
0
Such a solution exists if F does not grow too fast with respect to x, namely
kF (t, x)k = sup{kykX : y ∈ F (t, x)} ≤ c(t)(1 + kxkX ) on J × D,
with c ∈ L1 (J) and
F (t, x) ∩ TD (x) 6= ∅ on J × D,
where TD (x) is a tangent cone on X
{
}
−1
TD (x) = y ∈ X : lim λ ρ(x + λy, D) = 0 for x ∈ D.
λ→0+
We introduce the concept of measurability in multivalued maps
Definition 5. Let (D, A) be a measurable space, a multivalued map F : D → P(Rn )\{∅} is
called measurable if F −1 (B) ∈ A for all the open sets B ⊂ Rn .
There are some existence results of Cauchy problem (1.49). We present some of them. The basic
existence result for autonomous problems i.e. f independent of t is the following Lemma:
Lemma 9. [104, p. 53 Lem. 5.1] Let X = Rn , D ⊂ X be closed and F : D → P(Rn )\{∅} satisfy
(a) F is usc, F (x) is closed convex for all x ∈ D,
(b) kF (x)k ≤ c(t)(1 + kxkX ) on D, for some c > 0,
(c) F (x) ∩ TD (x) 6= ∅ on D.
Then (1.49) has a solution on R+ for every x0 ∈ D.
And the result for non-autonomous problems is:
Theorem 4 ([104, p. 58 Cor. 5.2]). Let X = Rn , J = [0, T ] ⊂ R, D ⊂ X be closed, and
F : J × D → P(X)\{∅} have closed convex values and be such that
(a) F (t, x) ∩ TD (x) 6= ∅ on J × D,
(b) kF (t, x)k ≤ c(t)(1 + kxkX ) on J × D, for c ∈ L1 (J),
(c) F (t, ·) is usc for all t ∈ J,
(d ) F (·, x) is measurable for all x ∈ D.
Then for any initial u0 ∈ D, the Cauchy problem (1.49) admits at least one absolutely continuous
solution on J.
Another important result
Corollary 1. [104, p. 59 Cor. 5.4] Let X = Rn , J = [0, T ] ⊂ R, D ⊂ X be closed, and
F : J × D → P(X)\{∅} have closed convex values and be such that
49
Chapter 1. Convergence of mass redistribution method for the one–dimensional
wave equation with a unilateral constraint at the boundary
(a) F (t, x) ∩ r(t, x)B1 (0) ∩ TD (x) 6= ∅ on J × D with r(t, x) = c(t)(1 + kxkX ). and c ∈ L1 (J),
(b) F (t, ·) is usc for all t ∈ J,
(c) F (·, x) is measurable for all x ∈ D.
Then for any initial u0 ∈ D, the Cauchy problem (1.49) admits at least one absolutely continuous
solution on J.
After proving the existence, uniqueness was investigated. Let X be a real Hilbert space with
inner product (·, ·) and F : J × D ⊂ X → P(X)\{∅} be monotone. If there exist k ∈ L1 (J) such
that
(1.50)
(
)
F (t, x) − F (t, y), x − y ≤ k(t)kx − yk2X , on J × D × D,
i.e. F be a semi–Lipschitz function. This ensures the uniqueness of the solution.
Theorem 5. [104, p. 138 Thm. 10.5] Let X be a real Hilbert space, J = [0, T ] ⊂ R, and
F : J × D → P(X)\{∅} have closed convex values and be such that
(a) F (t, ·) is usc,
(b) F (·, x) has a strongly measurable selection,
(
)
(c) condition F (t, x) − F (t, y), x − y ≤ ω(t, kx − ykX )kx − ykX holds with ω(t, ρ) = k(t)ρ and
k ∈ L1 (J),
(d ) kF (t, x)k ≤ c(t)(1 + kxkX ) on J × D, for c ∈ L1 (J),
are satisfied. Then Cauchy problem (1.49) has a unique solution on J.
Some functional results
We recall some functional results. First, we start with some definitions and results on semi–
groups. Let X be a Banach space with norm k · k. We denote by L(X) the Banach algebra of
all bounded linear operators on X endowed with the operator norm, still denoted by k · k.
Definition 6. A family (S(t))t>0 of bounded linear operators on a Banach space X is called a
(one–parameter) semi–group on X if it satisfies in the functional equation
(i) S(t + s) = S(t)S(s)
for all
t, s ≥ 0,
(ii) S(0) = I.
If (i) and (ii) hold even for all t, s ∈ R, we call (S(t))t∈R a (one–parameter) group on X.
Definition 7. A one–parameter semi–group (S(t))t>0 on a Banach space X is called uniformly
continuous (or norm continuous ) if
R+ 3 t 7→ S(t) ∈ L(X),
is continuous with respect to the uniform operator topology on L(X).
50
1.5. Numerical examples
The linear operator A defined by
{
}
S(t)x − x
D(A) = x ∈ X : lim
exists ,
t→0
t
and
S(t)x − x
d+ S(t)x
=
|t=0 for x ∈ D(A)
t→0
t
dt
Ax = lim
is the infinitesimal generator of the semi–group S(t). D(A) denotes the domain of A.
Theorem 6. [108] Every uniformly continuous semi–group (S(t))t>0 on a Banach space X is of
the form
S(t) = etA for all t > 0,
for some bounded operator A ∈ L(X).
Theorem 7. [109] A linear operator A is the infinitesimal generator of a uniformly continuous
semi–group if and only if A is a bounded linear operator.
Let H be a Hilbert space and for S ∈ L(H) denote by S ∗ its adjoint, i.e. the unique operator
satisfying (Sx | y) = (x | S ∗ y) for all x, y ∈ H. Now we take a uniformly continuous group
(S(t))t∈R on H and denote its generator by A, i.e.
S(t) = etA for all t ∈ R.
Since S 7→ S ∗ is continuous on L(H), it follows that the adjoint group (S(t)∗ )t∈R is again
uniformly continuous and that
∗
S(t)∗ = etA for all t ∈ R.
The groups for which all operators S(t) are unitary, i.e. satisfy (S(t))−1 = (S(t))∗ for all t ∈ R,
can be characterized as follows:
Proposition 2. The group (etA )t∈R is unitary if and only if A is skew–adjoint, i.e. A∗ = −A.
Definition 8. (S(t))t≥0 is a strongly continuous semi-group if the functional equation, defined
in Definition 6, holds and the orbit maps
ξx : t 7→ ξx (t) = S(t)x,
are continuous from R+ into X for every x ∈ X.
Proposition 3. For every strongly continuous semi–group S(t)t≥0 , there exist constant β ∈ R
and M ≥ 1 such that
kS(t)k 6 M eβt for all t ≥ 0.
Proposition 4. The adjoint semi–group of a strongly continuous semi–group on a reflexive
Banach space is again strongly continuous.
51
Chapter 1. Convergence of mass redistribution method for the one–dimensional
wave equation with a unilateral constraint at the boundary
Let X be a reflexive Banach space with norm k · k and dual X ∗ . Let A be an unbounded operator
in X with domain D(A) and let A be the generator of a strongly continuous semi–group S(t),
such that there exist constants M and β with
(i) kS(t)k 6 M eβt ,
(ii) k(A + λ)−k k 6 M (λ − β)−k , λ > β, k > 1.
The set of all operators A satisfying the conditions (i) and (ii) will be denoted by g(M, β). Let
{S(t)}t>0 be a unitary semi–group and kS(t)k 6 1, kS(t)k is called a contraction semi–group.
If A ∈ g(1, 0) then the operator −A is the infinitesimal generator of a semi–group which can be
written as S(t) = e−tA (see [110]).
For u0 ∈ D(A), there exists a unique function u in C1 ([0, T ]; X) ∩ C0 ([0, T ]; D(A)) such that
du
(t) + Au(t) = 0 for all t ∈ (0, T ),
dt
(1.51)
with initial condition
u(0) = u0 ,
then
u(t) = S(t)u0 .
This unique function u is called a strong solution. For more details we refer the reader to [22].
Let A ∈ g(M, β). The semi–group S(t) = e−tA gives the solution of the differential equation
(1.51) in the form u(t) = S(t)u(0). Let us consider the inhomogeneous differential equation for
f ∈ C0 (0, T ; X)

 du (t) + Au(t) = f (t)
for all
t ∈ (0, T ),
dt
(1.52)

u(0) = u0 .
The solution can be expressed with S(t) and depends linearly on u0 and f . We have
d
(S(t − s)u(s) = S(t − s)f (s))
ds
Integrating this expression over (0, t), we obtain

∫ t

 u(t) = S(t)u +
S(t − s)f (s) ds,
0
(1.53)
0

 u(0) = u .
0
In particular this implies that the solution of (1.52) is uniquely determined by u(0). Note that
if u is the unique solution to (1.52) and if f ∈ C0 ([0, T ]; X), then Au(t) ∈ C0 ([0, T ]; X) and the
equation (1.52) is verified in C0 ([0, T ]; X). For more details we refer the reader to [110]).
If S ∗ (t) is defined by
hS(t)u, u∗ i = hu, S ∗ (t)u∗ i for all u ∈ X and u∗ ∈ X ∗ ,
52
1.5. Numerical examples
then S ∗ (t) is a strongly continuous semi–group, whose generator A∗ is given by
D(A∗ ) = {u∗ ∈ X ∗ : u → hAu, u∗ i is continuous on D(A) in the norm topology of X},
and
hu, A∗ u∗ i = hAu, u∗ i for all u ∈ D(A) and u∗ ∈ D(A∗ ).
Definition 9. [105] A function u ∈ C0 ([0, T ]; X) is a weak solution of (1.52) if and only if for
every u∗ ∈ D(A∗ ) the function hu(t), u∗ i is absolutely continuous on [0, T ] and
d
hu(t), u∗ i = hu(t), A∗ u∗ i + hf (t), u∗ i,
dt
for almost all t ∈ [0, T ].
For the existence and uniqueness of this weak solution, we have the following Theorem:
Theorem 8. [105] For each u0 ∈ X there exists a unique weak solution u(t) of (1.52) satisfying
initial condition u(0) = u0 if and only if A is the generator of a strongly continuous semi–groups
S(t) of bounded linear operators on X, and in this case u(t) is given by (1.53).
Theorem 9. [Stone’s theorem [22]] Let A be a closed operator of domain D(A) dense in X, such
that A = iH with H self-adjoint, then A is the infinitesimal generator of a unitary group in X.
In the following Lemma, if u belongs to H2 and 2u is square integrable, then we define the traces
of u˙ on {t = 0} ∪ {t = T } and similarly traces u0 on {x = 0} ∪ {x = L}.
Lemma 10. Let u belongs to H2 , and assume that 2u = u
¨ − u00 (defined in the sense of distributions on QT ) is square integrable. Then, for ε > 0, we have
(1.54a)
(
)
u ∈ C0 [0, T ]; H1 (ε, L)
and
(
)
u˙ ∈ C0 [0, T ]; L2 (ε, L) ,
(1.54b)
(
)
u ∈ C0 [0, L]; H1 (ε, T − ε)
and
(
)
u0 ∈ C0 [0, L]; L2 (ε, T − ε) .
Proof. Let ψ belong to C∞ ([0, L]) be such that
{
1 on [ε, L],
ψ=
0 on [0, ε/2].
Let us define
u
˜ = uψ,
then we may deduce
2˜
u = ˜f ∈ L2 (QT ),
u
˜(0, t) = u
˜(L, t) = 0.
53
Chapter 1. Convergence of mass redistribution method for the one–dimensional
wave equation with a unilateral constraint at the boundary
If we let
X = H10 (0, L) × L2 (0, L)
and
D(A) = [H10 (0, L) ∩ H2 (0, L)] × H10 (0, L),
and with
[ ]
[ ]
u
v
A
= − 00
v
u
[ ]
u
for all û =
∈ D(A),
v
then A is skew–adjoint and therefore, with Stone’s theorem, A generates a unitary group S(t)
on X [89]. If we let
[ ]
[ ]
u
˜
0
û= ˙
and
F = ˜ ,
u
˜
f
then by Theorem 8, there is a unique weak solution û of dû
dt + Aû = F , and according to
Definition 9, and using the group property of S, û belongs to C0 ([0, T ]; X) which proves (1.54a).
To establish (1.54b), first let χ ∈ C∞ ([0, L]) be such that
{
1
on [ε, T − ε],
χ=
0
on [0, ε/2] ∪ [T − ε/2, T ].
Let us define
u
˜ = uχ.
Then, we can argue exactly as above.
54
Chapter 2
Numerical approximations of a one
dimensional elastodynamic contact
problem based on mass redistribution
method
This chapter is submitted for publication [111]
Contents
2.1
Introduction
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
56
2.2
Description of the models . . . . . . . . . . . . . . . . . . . . . . . . . . .
57
2.3
Discretization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
62
2.4
2.5
2.3.1
Classical finite element discretization . . . . . . . . . . . . . . . . . . . . . .
62
2.3.2
Finite element discretization with the mass redistribution method . . . . .
63
The wave equation with Signorini and Neumann boundary conditions
65
2.4.1
Analytical solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
65
2.4.2
Numerical simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
66
The wave equation with Signorini and Dirichlet boundary conditions .
69
2.5.1
Analytical solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
69
2.5.2
Numerical simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
70
2.5.2.1
The Crank–Nicolson method . . . . . . . . . . . . . . . . . . . . .
72
2.5.2.2
The Newmark method I . . . . . . . . . . . . . . . . . . . . . . . .
75
2.5.2.3
The Newmark method II . . . . . . . . . . . . . . . . . . . . . . .
76
2.5.2.4
The backward Euler method . . . . . . . . . . . . . . . . . . . . .
78
2.5.2.5
The Paoli–Schatzman method I . . . . . . . . . . . . . . . . . . .
79
2.5.2.6
The Paoli–Schatzman method II . . . . . . . . . . . . . . . . . . .
80
2.5.2.7
The Paoli–Schatzman method III . . . . . . . . . . . . . . . . . .
81
Appendix D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
82
Appendix E . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
84
Chapter 2. Numerical approximations of a one dimensional elastodynamic contact
problem based on mass redistribution method
2.1
Introduction
This paper aims to give some new numerical results for the dynamical evolution with impact of
a linearly elastic body and a rigid obstacle by using the mass redistribution method introduced
in [84]. The situations involving contact abound in industrial or biomedical problems. For this
reason a considerable engineering and mathematical literature has been devoted to contact problems. Usually, with the hypotheses of small deformations, the contact phenomena is modelled by
using the so–called Signorini’s boundary conditions in displacement, which are based on a linearization of the physically meaningful non penetrability of masses. The elastodynamic contact
problems are typically stated as hyperbolic type variational inequalities in Sobolev spaces, and
a few existence of solutions results has been established. The reader is referred to [40, 44, 45]
and also the comprehensive monograph [36]. The uniqueness result was successfully investigated for a wave equation in a half–space with a unilateral contact at the boundary in [44]. In
particular, the uniqueness still remains an open problem in other frameworks. The fundamental mathematical difficulties are related to the intrinsic non–smoothness of the problem coming
from the Signorini boundary conditions. Finally, notice that the existence of a strong solution
for viscoelastodynamic problems with unilateral constraints is proved in [53].
On the other hand, another challenging task consists to develop efficient numerical methods
for solving dynamical contact problems. Newmark’s method is one of the most popular numerical solvers which is usually combined with finite elements methods. Unfortunately, for almost
all values of the parameters of this method, unphysical energy blow–ups occur during the time
integration as well as numerical instabilities at the dynamical contact boundaries. To overcome
these difficulties, some numerical methods based on the Newmark scheme for solving impact
problems and which are energy dissipative have been introduced in [77]. Although stable, those
methods lead to an important energy loss at impact which do not vanish when the time step
decreases. Laursen et al. in [80, 78, 79, 112] have designed time integration schemes of Newmark
type which are energy conserving. However, these schemes are unable to prevent some spurious oscillations of the displacement and of the contact stress on the contact boundary. These
unphysical oscillations are avoided by the numerical methods developed in [113] (see also the
reference therein), but these methods are still energy dissipative. Another approach consisting
to remove the mass at the contact nodes is investigated in [84]. This mass redistribution method
prevents the oscillations at the contact boundaries mentioned above, and it can be proven that
the solution is unique and energy conserving, see [88]. Furthermore, some different variants has
been proposed in [114] with an estimate of the space approximation for linear elasticity. It has
also been extended for thin structures in [115, 116, 117] considering different discretizations for
the displacement and the velocity.
The paper is organised as follows. In Section 2.2, the mathematical models are presented and
a new proof of existence and uniqueness results is proposed. A space semi–discretization based
on a redistribution of mass is introduced in Section 2.3 and then some time integration method
are introduced. In Sections 2.4 and 2.5, two benchmark problems with their analytical solutions
are exhibited, one being new. More precisely, Section 2.4 is concerned by an elastodynamic
problem with Signorini and Neumann boundary while Section 2.5 deals with an elastodynamic
problem with Signorini and Dirichlet boundary conditions. Finally, numerical experiments for
56
2.2. Description of the models
these benchmark problems are reported and analysed. In particular, convergence rates for the
displacement, Lagrange multiplier and energy evolution show the efficiency of mass redistribution
method.
2.2
Description of the models
We consider a model for an elastic bar of length L vibrating longitudinally. One end of this bar
is free to move, as long as it does not hit a material obstacle, while the other end can be clamped
or free to move. The obstacle constrains the displacement of the extremity to be non negative.
The material of the bar is supposed to be homogeneous and the theory of small deformations
is considered. Let us denote by x the spatial coordinate along the bar, with the origin at the
material obstacle, let u(x, t) be the displacement at time t ∈ [0, T ], T > 0 of the material point of
spatial coordinate x ∈ [0, L]. Let f (x, t) denotes a density of external forces, depending on time
and space. In the case where the bar is clamped at one end (see Figure 2.5), the mathematical
problem is formulated as follows:
utt (x, t) − uxx (x, t) = f (x, t),
(2.1)
(x, t) ∈ (0, L) × (0, T ),
with Cauchy initial data
u(x, 0) = u0 (x) and ut (x, 0) = v 0 (x),
(2.2)
x ∈ (0, L),
and Signorini and Dirichlet boundary conditions at x = 0 and x = L, respectively,
0 ≤ u(0, t) ⊥ ux (0, t) ≤ 0 and u(L, t) = 0,
(2.3)
t ∈ [0, T ].
While in the case where the bar is free to move at one end (see Figure 2.1), the motion is
governed by the momentum equilibrium equation (2.1) together with Cauchy initial data (2.2)
and Signorini and Neumann boundary conditions at x = 0 and x = L, respectively,
0 ≤ u(0, t) ⊥ ux (0, t) ≤ 0 and ux (L, t) = 0,
(2.4)
def
t ∈ [0, T ].
def
∂u
Here ut = ∂u
∂t and ux = ∂x . The orthogonality has the natural meaning; namely if we have
enough regularity, it means that at the boundary x = 0 the product u(0, ·)ux (0, ·) vanishes almost
everywhere in time. If it is not the case, the above inequality is integrated on an appropriate set
of test functions, leading to a weak formulation for the unilateral condition. Observe that from
mathematical viewpoint, Signorini’s condition means that when the bar touches the obstacle at
x = 0, its reaction can be only upwards, so that ux (0, ·) ≤ 0 on the set {t : u(0, ·) = 0}. While in
the case where the bar does not touch the obstacle, its end is free to move. More precisely, we
have ux (0, ·) = 0 on the set {t : u(0, ·) > 0}. We suppose that the initial displacement u0 belongs
to the H1 (0, L) and satisfies the compatibility conditions, i.e. u0 (L) = 0 and u0 (0) ≥ 0, the initial
velocity v 0 belongs to L2 (0, L) and the density of external forces f belongs to L2 (0, T ; L2 (Ω)).
We describe now the weak formulations of both the two problems. For that purpose, it is
def
def
convenient to introduce the following notations: VDir = {u ∈ H1 (0, L) : u(L) = 0}, VNeu =
def
H1 (0, L) and H = L2 (0, L). We denote by K the following convex set:
K = {u ∈ H2 : u(·, t) ∈ K for almost every t},
def
57
Chapter 2. Numerical approximations of a one dimensional elastodynamic contact
problem based on mass redistribution method
where H2 = {u ∈ L2 (0, T ; V) : ut ∈ L2 (0, T ; H)}, K = {u ∈ V : u(0) ≥ 0} and V = VDir
or V = VNeu depending on the conditions taken in x = L. Therefore the weak formulations
associated to (2.1)–(2.3) and to (2.1)–(2.2) together with (2.4) are obtained by multiplying (2.1)
def
by v − u and by integrating formally this result over QT = (0, L) × (0, T ) to get
def
(2.5)
def


find u ∈ K such that


∫
∫
∫
 ∫ L
0
0
−
v (v(·, 0)−u ) dx−
ut (vt −ut ) dx dt+
ux (vx −ux ) dx dt≥
f (v−u) dx dt

0
QT
QT
QT



for all v ∈ K for which there exists ζ > 0 with v = u for t ≥ T − ζ.
Notice that the sole difference of the weak formulation associated to (2.1)–(2.3) and to (2.1)–
(2.2), (2.4) comes from the definition of the convex set K.
It should also be noted that existence and uniqueness results are obtained for a similar
situation of a vibrating string with concave obstacle in one dimensional space in [40] and also for
a wave equations with unilateral constraint at the boundary in a half space of RN in [44]. An
existence result for a wave equation in a C2 –regular bounded domain constrained by an obstacle
at the boundary in RN for N ≥ 2 is proven in [45].
Theorem 10 (Existence and uniqueness results). Assume that u0 ∈ V, v 0 ∈ H and f ∈
L2 (0, T ; H). Then Problem (2.1)–(2.3) admits a unique solution u ∈ L∞ (0, T ; V)∩W1,∞ (0, T ; H).
Proof. The first step consists to introduce an auxiliary problem. Namely, let u
¯ be the solution to Problem (2.1) with initial data (2.2) and boundary conditions (2.3) where Signorini’s
boundary condition is replaced by Dirichlet’s one, i.e. u
¯(0, t) = 0, t ∈ [0, T ]. Note that the
auxiliary problem possesses a unique solution u
¯ belonging to C0 ([0, T ]; H2 (0, L) ∩ H10 (0, L)) ∩
def
¯ a solution of
C1 ([0, T ]; H10 (0, L)) ∩ C2 ([0, T ]; H) (see [118, 119]). Then we denote by v = u − u
vtt (x, t) − vxx (x, t) = 0,
(2.6)
(x, t) ∈ (0, L) × (0, T ),
with Cauchy initial data
(2.7)
v(x, 0) = 0 and vt (x, 0) = 0,
x ∈ (0, L),
and Signorini and Dirichlet boundary conditions at x = 0 and x = L, respectively,
0 ≤ v(0, t) ⊥ vx (0, t) + u
¯x (0, t) ≤ 0 and v(L, t) = 0,
(2.8)
t ∈ [0, T ].
def
Define g(t) = u
¯x (0, t). The second step consists to rewrite (2.6)–(2.8) as a differential inclusion
problem by using the characteristic method. To this aim, it is convenient to introduce the
characteristic coordinates
def
def
ξ = x + t and η = x − t.
Therefore the chain rule gives
∂2v
∂2v
∂2v
∂2v
=
+
2
+
∂x2
∂ξ 2
∂ξ∂η ∂η 2
and
∂2v
∂2v
∂2v
∂2v
=
−
2
+
.
∂t2
∂ξ 2
∂ξ∂η ∂η 2
2
∂ v
According to (2.1), ∂ξ∂η
vanishes and it follows that v(ξ, η) = φ(ξ) + ψ(η), where φ and ψ are
two differentiable functions such that
(2.9)
v(x, t) = φ(x+t) + ψ(x−t).
58
2.2. Description of the models
In particular, taking t = 0 and using the Cauchy initial data, we get
(2.10)
φ(x) + ψ(x) = 0 and φ0 (x) − ψ 0 (x) = 0,
which by integration implies that there exist two constants C φ and C ψ such that φ(ξ) = C φ and
ψ(η) = C ψ for all ξ and η belonging to [0, L]. According to (2.9), the boundary conditions (2.3)
can be rewritten as follows:
(2.11)
0 ≤ φ(t) + ψ(−t) ⊥ φ0 (t) + ψ 0 (−t) + g(t) ≤ 0 and φ(L+t) + ψ(L−t) = 0
for all t belonging to [0, T ]. Thanks to the above identity, we may extend φ(t) for all t ∈ [L, 2L],
i.e. we have
φ(L+t) = −ψ(L−t)
for all t belonging to [0, L]. If we choose t0 = L + t, we get φ(t0 ) = −ψ(2L−t0 ). We already have
the solution for ψ(t) with 0 ≤ t ≤ L and if L ≤ t0 ≤ 2L, we can obtain φ(t0 ) by observing that
0 ≤ 2L − t0 ≤ L and by using ψ(2L−t), it comes that φ(t) = C ψ for all t belonging to [L, 2L].
¯ defined by
We introduce the multivalued function JN : R → P(R)


if x < 0,

{0}
def
JN [x] = [0, +∞) if x = 0,


∅
if x > 0,
¯ is the set of all subsets of R.
¯ More precisely, JN (x) is the subdifferential of the
where P(R)
indicator function defined by
{
if x ∈ (−∞, 0],
def 0
I(−∞,0] (x) =
+∞ if x ∈
/ (−∞, 0].
Obviously, I(−∞,0] is a lower semi–continuous and convex function, for further details the reader
is referred to [91]. Then, the inequalities in (2.11) can be rewritten as follows
(2.12)
φ0 (t) + ψ 0 (−t) + g(t) ∈ −JN [−φ(t)−ψ(−t)].
Note that at this stage, ψ(η), η ∈ [−L, L], is the unique unknown of (2.12). Let us define now
def
w(t) = −φ(t) − ψ(−t).
(2.13)
We insert (2.13) into (2.12) to get
w0 (t) ∈ −JN [w(t)] − g(t).
Finally, we find the following Cauchy problem
(2.14a)
w0 (t) ∈ −JN [w(t)] − g(t) a.e. t ∈ (0, L),
(2.14b)
w(0) = 0.
59
Chapter 2. Numerical approximations of a one dimensional elastodynamic contact
problem based on mass redistribution method
Then we look for a solution of the following problem:
(2.15a)
w0 (t) + h(t) = −g(t) a.e. t ∈ (0, L),
(2.15b)
h(t) ∈ JN [w(t)] a.e. t ∈ (0, L),
(2.15c)
w(0) = 0.
We observe that (2.15) is equivalent to (2.11), the verification is left to the reader. We regularize
the problem (2.15), we construct a solution of the regularized problem, and we show the existence
of a solution by passing to the limit with respect to regularity parameter.
We regularize the problem (2.15) by using Yosida’s regularization. More precisely, we introduce first a sequence {gε }ε>0 such that gε belongs to C0 ([0, L]) for all ε > 0 and
gε → g
(2.16)
in H as ε & 0.
We may approximate (2.15) by the following problem:
(2.17a)
wε0 (t) + JN,ε [wε (t)] = −gε (t) a.e. t ∈ (0, L),
(2.17b)
wε (0) = 0.
where JN,ε = 1ε (1−Jε ) denotes the Yosida approximation with the resolvant defined by Jε =
(1+εJN )−1 . Note that JN,ε is a monotone and Lipschitz continuous mapping defined on all H.
Clearly the contraction principle step by step in time allows us to prove that (2.15) possesses a
unique solution wε belonging to C1 ([0, L]).
def
def
We establish now some a priori estimates, which later will enable to infer the existence of a
weak solution. To this aim, we multiply (2.17a) by wε0 (t), and we integrate this expression over
(0, L) to get
∫ L
∫ L
∫ L
0
0
2
gε (t)wε0 (t) dt.
|wε (t)| dt +
JN,ε [wε (t)]wε (t) dt = −
(2.18)
0
0
0
Let us define now
def
ε
I(−∞,0]
(w) = min
z∈H
{1
}
2
|w−z|
+
I
(z)
(−∞,0]
2ε
for all ε > 0 and w ∈ H. Since JN is the subdifferential of a proper, convex and lower semiε
continuous function I(−∞,0] , it follows that I(−∞,0]
is convex, Fréchet differentiable in H and its
ε
subdifferential ∂I(−∞,0] coincides with JN,ε . Furthermore we have
ε
∀z ∈ H : I(−∞,0]
(z) % I(−∞,0] (z) as ε & 0.
Then we may deduce from (2.18) that
(2.19)
1
1 0 2
ε
kwε kH + I(−∞,0]
(wε (L)) ≤ kgε k2H .
2
2
ε
We observe that by definition I(−∞,0]
(wε (L)) ≥ 0 which follows from (2.19) that there exists
C1 > 0, independent of ε > 0, such that
(2.20)
kwε0 kH ≤ C1 .
60
2.2. Description of the models
Furthermore we have
|wε (τ )|2 ≤ 2Lkwε0 k2H
(2.21)
for all τ ∈ (0, L). Then we deduce from (2.19) and (2.21) that there exists C2 > 0, independent
of ε > 0, such that
|wε (τ )|2 + kwε0 k2H ≤ C2
(2.22)
for all τ ∈ (0, L). Since Jε is a contraction on all H (see [91]), it follows from (2.22) that there
exists C3 > 0, independent of ε > 0, such that
kwε kV + kJε wε kV ≤ C3 .
(2.23)
We observe that (2.17a) implies that
kJN,ε [wε ]k2H ≤ 2kwε0 k2H + 2kgε k2H .
(2.24)
The above a priori estimates allow us to infer that there exists w and h belonging to V∩L∞ (0, L),
respectively, and passing to subsequences, if necessary, we find
Jε wε * w
(2.25a)
(2.25b)
in L∞ (0, L) weak ∗,
JN,ε [wε ] * h in H weak,
as ε tends to 0. Since V ,→ H with compact embedding, it follows that
(2.26a)
Jε wε → w
(2.26b)
wε → w
in H,
in H and wε * w
in V
weak,
as ε tends to 0. It remains to prove that w and h satisfy (2.15). We integrate (2.17a) over (0, τ ),
we get
∫ τ
( 0
)
(2.27)
wε (t)+JN,ε [wε (t)]+gε (t) dt = 0
0
for all τ ∈ (0, L). Thanks to (2.25) and (2.26), we may pass to the limit in all the terms of (2.27),
namely we find
∫ τ
( 0
)
w (t)+h(t)+g(t) dt = 0
0
for all τ ∈ (0, L), which is equivalent to (2.15a). Let us note that JN,ε [wε ] ∈ JN [Jε wε ], for some
technical details, the reader is referred to [91]. Then (2.25b) and (2.26a) enable us to deduce
that
∫ L
∫ L
(2.28)
lim sup
JN,ε [wε (t)]Jε wε (t) dt ≤
h(t)w(t) dt,
ε&0
0
0
which implies that (2.15b) holds. Then we may conclude that there exists a solution v to Problem
(2.6)–(2.8).
61
Chapter 2. Numerical approximations of a one dimensional elastodynamic contact
problem based on mass redistribution method
The uniqueness result comes from the monotonicity of JN , namely JN is the subdifferential
of a convex, lower semi–continuous and proper function (see [91] for further details).
The existence and uniqueness results to Problem (2.1)–(2.2) together with (2.4) is established
by using analogous approach developed above for Problem (2.1)–(2.3). The main difference comes
from the choice of the auxiliary problem. Let u
¯ be the unique solution to Problem (2.1) with
initial data (2.2) and boundary conditions (2.4) where the Signorini boundary conditions are
replaced by Dirichlet boundary conditions, i.e. u
¯(0, t) = 0, t ∈ [0, T ]. Hence the rest of the proof
is quite a routine, the verification is left to the reader.
Finally, we consider the following energy which is associated to (2.5):
1
E(t) =
2
∫
L(
)
|ut (x, t)|2 +|ux (x, t)|2 dx.
0
This energy is constant with respect to time t when the density forces f vanishes.
Note that formulations (2.5) and (2.15) are equivalent (in the sense that they have the same
weak solutions), since the proof given in [88] can be straightforwardly adapted.
2.3
2.3.1
Discretization
Classical finite element discretization
We introduce now a semi–discrete problem in space associated to (2.5) . To this aim, we choose
def L
a set of parameters h = m
(mesh size) having in mind the limit m → +∞, where m is an integer,
def
def
Neu
0
and let Vh = {vh ∈ C ([0, L]) : vh |[ai ,ai+1 ] ∈ P1 , i = 0, . . . , m − 1} and VhDir = {vh ∈ VhNeu :
def
vh (L) = 0}. Here, ai = ih, i = 0, 1, . . . , m and P1 is the space of polynomials of degree less than
or equal to 1. A classical basis of Vh is given by the sequence of shape functions ϕi ∈ Vh for
i = 0, 1, . . . , J, defined by
{
def
ϕi (x) =
1−
0
|x−ai |
h
if
x ∈ [amax(i−1,0) , amin(i+1,m) ],
otherwise,
where J = m − 1 in the case where Vh = VhDir and J = m in the case where Vh = VhNeu . Observe
that ϕi (aj ) = δij , j = 0, 1, . . . , m (δ is Kronecker’s symbol). We approximate the solution u
belonging to V to the weak formulation (2.5) by
uh (x, t) =
J
∑
uj (t)ϕj (x).
j=0
Consequently, we have ui = uh (ai ), i = 0, 1, . . . , J. We shall consider two strategies to approximate the dynamic contact problem. The first one is a classical finite element semi–discretization
62
2.3. Discretization
which uses a multiplier. It reads as follows:


find uh : [0, T ] → Vh and λ : [0, T ] → R such that for all vh ∈ Vh


∫ L
∫ L
∫ L




uh,tt vh dx +
uh,x vh,x dx = −λvh (0) +
f vh dx a.e. t ∈ [0, T ],
(2.29)
0
0
0



0 ≤ uh (0, ·) ⊥ λ ≤ 0 a.e. t ∈ [0, T ],



u (·, 0) = u0 and u (·, 0) = v 0 ,
h
h
h,t
h
where u0h and vh0 belong to Vh and λ is the Lagrange multiplier being here the contact force.
Furthermore, the approximated problem (2.29) can be rewritten in an algebraic formulation form
as follows:

n

find Uh : [0, T ] → R and λ : [0, T ] → R such that


M U + SU = −λe + F a.e. t ∈ [0, T ],
0
h,tt
h
(2.30)

0 ≤ u0 ⊥ λ ≤ 0 a.e. t ∈ [0, T ],




Uh (0) = Uh0 and Uh,t (0) = Vh0 ,
def
where M and S denote the mass
and stiffness
matrices, respectively, and U = (u0 , . . . , un−1 )T ,
∫
∫
def
def
L
L
e0 = (1, 0, . . . , 0)T and F = ( 0 f ϕ0 dx, . . . , 0 f ϕn−1 dx).
2.3.2
Finite element discretization with the mass redistribution method
An alternative to the standard discretization is to consider a trivial mass redistribution method
which consists to replace the mass matrix M in (2.30) by a modified mass matrix M mod defined
by
∫ L
mod def
ϕi ϕj dx for all i, j ∈ [0, n − 1].
Mij =
h
Let us observe that there are more sophisticated mass redistribution methods, see [90]. The
modified mass matrix has the form
(
)
0 0
mod def
M
=
,
¯
0 M
¯ admits an unmodified part M
¯ ij = Mi+1,j+1 , for all i, j = 1, . . . , n − 2. As a consequence
where M
and in a certain sense, the node at the contact boundary evolves in a quasi–static way. It has
been established in [88] in a slightly different context that the approximated solution using the
mass redistribution method converges to the unique solution to (2.5).
Finally, note that the discrete energy associated to problem (2.30) is given by
(2.31)
1 T
T
Eh (t) = (Uh,t
M Uh,t +UhT SUh −Uh,t
F )(t).
2
We introduce now the time discretization. In order to fix the notations, let the time interval [0, T ]
be divided by n + 1 discrete time–points such that 0 = t0 < t1 < . . . < tn = T . Furthermore the
n , U n and λn are assumed to be given by algorithmic approximations
discrete quantities Uhn , Uh,t
h,tt
63
Chapter 2. Numerical approximations of a one dimensional elastodynamic contact
problem based on mass redistribution method
of the displacement Uh (tn ), the velocity Uh,t (tn ) and the acceleration Uh,tt (tn ) and the Lagrange
multiplier λ(tn ), respectively. Some time–stepping schemes allowing to obtain an approximated
solution to Problem (2.1)–(2.4) are introduced below and their efficiency is discussed and analyzed
in the next sections.
The most wide–spread time-stepping scheme for solving contact problems is the family of
classical Newmark methods proposed by Newmark in [74]. The algorithms form a subset of the
Hilber–Hughes–Taylor (HHT) family of temporal integrators, the reader is referred to [75] for
further details. The underlying concept of the discretizations are Taylor expansions of displacements and velocities neglecting terms of higher order. For the contact problem introduced above,
the discrete evolution is described by the finite difference equations:
(1
)
n+1
n
−β ∆t2 Uh,tt
,
+ β∆t2 Uh,tt
(2.32a)
n
Uhn+1 = Uhn + ∆tUh,t
+
(2.32b)
n+1
n+1
n
n
Uh,t
= Uh,t
+ (1−γ)∆tUh,tt
+ γ∆tUh,tt
,
n+1
n+1
n+1
n+1
M Uh,tt + SUh = −λ
e0 + F
,
0 ≤ un+1
⊥ λn+1 ≤ 0,
0
(2.32c)
(2.32d)
2
where ∆t is a given time step and (β, γ) are the algorithmic parameters. Furthermore Uh0 ,
0 and λ0 are given and U 0
Uh,t
h,tt is evaluated by using (2.32c). Notice that the Newmark family
contains many well–known and widely–used algorithms which correspond to different choices of
the algorithmic parameters (β, γ) (see [120, 112]). This method is unconditionally stable for
linear elastodynamic problem for γ ≥ 12 and β ≥ 14 ( 21 +γ)2 (see [120, 121]). One of the most
used method to study dynamic problems is the Crank–Nicolson method, also called trapezoidal
or average acceleration method, which is obtained by setting the parameters (β, γ) = ( 41 , 12 ).
This method is second–order consistent and unconditionally stable in the unconstrained case.
Furthermore, the total energy of the discrete evolution is preserved, for the purely elastic case, see
[112]. That is the reason why the Crank–Nicolson method is commonly used in the community
of computational mechanics. However, the situation should be examined in the case of contact
constraints. Indeed the order of accuracy is degraded, for further details, the reader is referred
def
to [120, 121, 122]. Let us define now the energy evolution by ∆Ehn = Ehn+1 − Ehn , where Ehn is
assumed to be given by an algorithmic approximation of the energy Eh (tn ). In particular, the
energy evolution associated to (2.32) for (β, γ) = ( 14 , 12 ) is given by
∆Ehn = −(Vhn )T Λnγ e0 ,
(2.33)
where Vhn = Uhn+1 − Uhn and Λnγ = (1−γ)λn + γλn+1 . Let us observe that in the presence
of permanent contact, the energy is strictly conserved, while the release of an existing contact
decreases the energy and the detection of a new contact increases the energy. However, regain
and loss of energy do not balance and there is no guarantee that the energy will stay bounded
during the time integration.
def
def
Another approach to study elastodynamic contact problems consists to use the θ–method.
64
2.4. The wave equation with Signorini and Neumann boundary conditions
More precisely, the discrete evolution is described by the finite difference equations:
(2.34a)
n+1
n
Uhn+1 = Uhn + ∆t((1−θ)Uh,t
+θUh,t
),
(2.34b)
n+1
n+1
n
n
Uh,t
= Uh,t
+ ∆t((1−θ)Uh,tt
+θUh,tt
),
(2.34c)
n+1
M Uh,tt
+ SUhn+1 = −λn+1 e0 + F n+1 ,
(2.34d)
0 ≤ un+1
⊥ λn+1 ≤ 0,
0
0 and λ0 are given and U 0
with θ ∈ [0, 1]. Furthermore Uh0 , Uh,t
h,tt is evaluated by using (2.34c).
The energy evolution associated to (2.34) is given by
(1 )
(1 )
n T
n
(2.35)
∆Ehn =
−θ (Vh,t
) M Vh,t
+
−θ (Vhn )T SVhn − (Vhn )T Λnθ e0 .
2
2
Notice that (2.34) with θ = 0 and θ = 1 are called the forward and backward Euler method,
respectively, while with θ = 12 , we obtain Crank–Nicolson’s method.
Finally, we focus on the so–called Paoli–Schatzman method that consists to fix the contact
constraint at an intermediate time step. Indeed the method proposed below is a slight modification of Paoli–Schatzman method (see [85, 123]) which takes into account the kernel of the
modified mass matrix. A simple application of Paoli–Schatzman method based on Newmark
scheme to Problem (2.30) with γ = 21 leads to


find Uhn+1 : [0, T ] → Rn and λn : [0, T ] → R such that:





M (Uhn+1 −2Uhn +Uhn−1 )


+ S(βUhn+1 +(1−2β)Uhn +βUhn−1 ) = −λn e0 for all n ≥ 2,
2
∆t
(2.36)
un+1
+ eun−1

0
0

0 ≤ un,e
⊥ λn ≤ 0,
=

0

1
+
e


U and U given.
0
1
Here e belongs to [0, 1] and is aimed to be interpreted as a restitution coefficient. Note that Uh0
and Uh1 are given data and Uh1 can be evaluated by a one step scheme. We may observe that
taking M = M mod in (2.36), we are not able to resolve the problem on the kernel of M mod . That
is the reason why, SUhn−1 as well as SUhn are projected on the orthogonal of the kernel of M .
The energy evolution associated to (2.36) is given for β = 14 by
1+e n n−1
λ u0 .
2
Notice that the numerical simulations were performed by employing the finite element library
Getfem++ (see [87]). For the reader convenience, the energy evolutions (2.33), (2.35) and (2.37)
are justified in the Appendix.
∆Ehn =
(2.37)
2.4
2.4.1
The wave equation with Signorini and Neumann boundary
conditions
Analytical solution
We consider in this section the analytical solution proposed in [103]. It is the solution to the
equation of motion (2.1) together with Cauchy initial data (2.2) and boundary conditions (2.4) in
65
Chapter 2. Numerical approximations of a one dimensional elastodynamic contact
problem based on mass redistribution method
the case where the density of external forces f (x, t) = −(g+ux (0, t)) with g > 0 being the gravity
acceleration. Let us describe the solution of this problem (see Figure 2.1). Before the impact,
the bar is undeformed
and its initial velocity v 0 vanishes. The bar reaches the rigid obstacle at
√
√
2u0
time t1 =
2u0 g. After the impact, the bar stays in contact as long
g with the velocity
as the shock wave travels from bottom to the top of the bar and vice versa and then it takes
off. Since the velocity of the shock wave c is assumed to be equal to 1, the bar stays in contact
with the rigid obstacle a time L. Finally, notice that impacts occur at t4k+1 = 3L + 16kL,
t4k+2 = t4k+1 + 2L, t4k+3 = t4k+1 + 8L and t4k+4 = t4k+1 + 10L. We may deduce that the
L
x=0
t=0
Figure 2.1: An elastic bar with both ends free to move
analytical solution is given by

u0 − 12 gt2






h (x, t−t4k+1 )

 1
u(x, t) = h2 (x, t−t4k+2 )



h1 (x, t4k+4 −t)

(


u0 − 1 g t−t
(2.38)
2
√ )2
2u0
−
4k+4
g
if
t ≤ t1 ,
if
t4k+1 < t ≤ t4k+2 ,
if
t4k+2 < t ≤ t4k+3 ,
if
t4k+3 < t ≤ t4k+4 ,
if
t4k+4 < t ≤ t4(k+1)+1 .
where we have used the following notations
√
+∞
∑
2g
2u0
def
(2.39)
h1 (x, t) = −
min(x, L−|t−L|) +
(cos(νn t)−1) sin(νn x),
g
Lνn3
n=1
√
+∞
(
1
2u0 )2 2gL2 ∑ 4g
def
(2.40)
h2 (x, t) = u0 − g t−
−
+
cos(λn t) cos(λn x),
2
g
3
λ2n
n=1
def
def
with νn = (n− 12 ) Lπ and λn = n Lπ .
2.4.2
Numerical simulations
The parameters used in the numerical simulations are L = 1, g = 0.11, u0 (x) = 0.5, v 0 (x) = 0
and ux (L, ·) = 0.
66
2.4. The wave equation with Signorini and Neumann boundary conditions
Figure 2.2: First analytical solution: comparison of the analytical solution (u, λ) and approximated solutions (unh , λn ) by using standard mass matrix (left) and modified mass matrix (right)
in the contact.
We evaluate the approximated solution (Uhn , λn ) to Problem (2.1)–(2.3) by using Crank–Nicolson
method that consists to choose (β, γ) = ( 14 , 21 ) in (2.32) for the standard and modified mass
matrices. Then we compare the solutions obtained by employing Crank–Nicolson method to
the analytical one (u, λ) in different points of the bar (see Figure 2.2). In the case where the
constraint is active, the numerical experiments highlighted some spurious oscillations occurring
for the approximated solution (Uhn , λn ) obtained by solving (2.32) with the standard mass matrix
M (see Figure 2.2 (left)) while these oscillations disappeared when the standard mass matrix is
replaced by the modified mass matrix, i.e. M = M mod (see Figure 2.2 (right)). However the
approximated solutions obtained by both methods converge to the analytical solution when both
the space ∆x and time ∆t steps tend to 0 in Lp (0, T ; H) norms with p = 2, +∞, see Figure 2.3
(a constant ratio ∆x
∆t = 1 has been used).
Surprisingly, the error curves in the norms kuh −ukLp (0,T ;V) , with p = 2, +∞ shown in the same
figure are both diverging for the standard and modified mass matrices. The explanation of this
67
Chapter 2. Numerical approximations of a one dimensional elastodynamic contact
problem based on mass redistribution method
(a) kuh −ukL∞ (0,T ;H)
(b) kuh −ukL2 (0,T ;H)
(c) kuh −ukL∞ (0,T ;V)
(d) kuh −ukL2 (0,T ;V)
Figure 2.3: First analytical solution: comparison of the error curves obtained by using standard
and modified mass matrices. A constant ratio ∆x
∆t = 1 has been used
Figure 2.4: First analytical solution: representation of the derivative of displacement with respect
to x of analytical (left) and approximated (right) solutions. The analytical solution is computed
with the 103 first terms of the series.
68
2.5. The wave equation with Signorini and Dirichlet boundary conditions
divergence is illustrated in Figure 2.4 where the space derivative of the displacement is shown for
both the analytical and the approximated solution. Indeed, Figure 2.4 (left) illustrates the fact
that the series of the derivative corresponding to (2.40) do not converge. This is why it seems
not to be possible to use this analytical solution to estimate the error in Lp (0, T ; V) norm. For
that reason, another analytical solution is exhibited in the next section.
2.5
2.5.1
The wave equation with Signorini and Dirichlet boundary
conditions
Analytical solution
In this section, we propose a new analytical piecewise affine and periodic solution to the problem
composed by the momentum equilibrium equation (2.1) together with Cauchy initial data (2.2)
and boundary conditions (2.3) in the case where the density of external forces f (x, t) vanishes.
Suppose that the initial displacement and velocity are given by u0 (x) = 12 − x2 and v 0 (x) = 0,
respectively. The length of bar is L = 1, namely, the bar is compressed at t = 0. Let us describe
the solution of this problem (see Figure 2.5). Before the impact, the bar, which is clamped at
one end, elongates until it reaches the rigid obstacle at t1 = 1. After the impact, the bar stays
in contact for the time t ∈ (t1 , t2 ) and it takes off at time t2 = 2.
L
x=0
t=0
t1 = 1
t2 = 2
t3 = 3
Figure 2.5: An elastic bar with one end fixed and the other end free to move
We use characteristics method to determine the solution of the above problem. To this aim, we
split each domain defined by (0, L) × (ti , ti+1 ), i = 0, 1, 2, corresponding to the phases before
the impact, during the impact and after the impact, respectively, into four regions according to
Figure 2.6. We choose below L = 1 and ti = i, i = 0, . . . , 3.
def
def
Before the impact, according to characteristics lines ξ = x+t and η = x−t, we divide the domain
(0, 1) × (0, 1) into four regions as it is drawn on Figure 2.6. Then we observe that ξ ∈ [0, 1] and
η ∈ [0, 1] in the region I, ξ ∈ [0, 1] and η ∈ [−1, 0] in the region II, ξ ∈ [1, 2] and η ∈ [0, 1] in the
region III and ξ ∈ [1, 2] and η ∈ [−1, 0] in the region IV. Therefore we may conclude that the
solution is given by
(2.41a)
u(x, t) = 12 (1−x) in the regions I, III,
(2.41b)
u(x, t) = 21 (1−t) in the regions II, IV.
69
Chapter 2. Numerical approximations of a one dimensional elastodynamic contact
problem based on mass redistribution method
x
before
contact
L=1
during
contact
II
I
after
contact
II
III
IV
I
II
III
I
IV
III
IV
t
0
t0 = 0
u = 1−x
2
v=0
t1 = 1
u=0
v = −1/2
t2 = 2
u=0
v = 1/2
t3 = 3
u = 1−x
2
v=0
Figure 2.6: The regions allowing to determine the value of u.
Once again by using the characteristics lines, we divide the domain (1, 2)×(0, 1) into four regions.
Then the solution (2.41b) calculated in the region IV allows us to evaluate the initial conditions
at time t1 = 1, namely we found u(·, 1) = 0 and ut (·, 1) = − 12 . Therefore we obtain
(2.42a)
u(x, t) = − 12 (t1 −t) in the region I,
(2.42b)
u(x, t) = − x2 in the region II,
(2.42c)
u(x, t) = 12 (x−1) in the region III,
(2.42d)
u(x, t) = 12 (t−t1 −1) in the region IV.
After the impact, we divide the domain (2, 3) × (0, 1) into four regions. It follows from (2.42d)
that u(·, 2) = 0 and ut (·, 2) = 12 which implies that
u(x, t) = 12 (t−t2 ) in the regions I, II,
u(x, t) = 12 (1−x) in the regions III, IV.
Furthermore we observe that u(·, 3) = u0 and ut (·, 3) = v 0 . Then we conclude that the solution
is periodic of period 3. Note that the Lagrange multiplier λ is equal to the contact force ux (0, ·).
2.5.2
Numerical simulations
We evaluate below (Uhn , λn ) by using different time–stepping methods, namely the Newmark
method (2.32), the backward Euler method (2.34) with θ = 1 and the Paoli–Schatzman method
(2.36). More precisely, for each time–stepping method, the approximated solutions (Uhn , λn )
are obtained by solving (2.32) for different values of the couple (β, γ), (2.36) for β = 14 and
different values for the parameter e and (2.34) in the particular case where θ = 1, respectively.
The computation are also performed for both the standard mass matrix M and the modified
one M = M mod and they are compared to the analytical solution (u, λ) exhibited in Section
2.5.1 (see Figures 2.7, 2.11, 2.13, 2.15, 2.17, 2.19 and 2.21). Furthermore, the energy evolution
associated to different methods mentioned above are presented in Figures 2.8, 2.12, 2.14, 2.16,
2.18, 2.20 and 2.22. Note that the numerical experiments were still done for different space and
time steps with a constant ratio ∆x
∆t = 10. Some error curves for Crank–Nichoslon’s method are
70
2.5. The wave equation with Signorini and Dirichlet boundary conditions
presented for standard matrix as well as for modified mass matrix. Notice that the slope of error
curves for nonlinear problems reflect both the regularity of solution and the ability of numerical
schemes to reproduce some technical inequalities.
Scheme
Crank–Nicolson
Method
kuh −ukL∞ (0,T ;H)
kuh −ukL2 (0,T ;H)
kuh −ukL∞ (0,T ;V)
kuh −ukL2 (0,T ;V)
STD
0.48045
0.47113
-0.30273
-0.31299
MOD
0.88075
0.97113
0.38624
0.36192
Newmark I
STD
0.51444
0.51768
0.26104
0.24000
(β, γ) = ( 12 , 1)
MOD
0.49236
0.48034
0.28219
0.25943
Newmark II
STD
0.57503
0.61856
0.34411
0.30100
MOD
0.95055
1.00090
0.39734
0.33789
STD
0.51804
0.51771
0.26784
0.24847
MOD
0.50425
0.49024
0.27753
0.25465
Paoli–Schatzman I
STD
0.66064
0.66015
0.31204
0.28406
e=0
MOD
0.98009
1.00690
0.39570
0.41096
(β, γ) =
( 12 , 12 )
Backward Euler
Paoli–Schatzman II
STD
0.66064
0.66014
0.33914
0.30116
MOD
0.97659
1.00170
0.41159
0.42823
Paoli–Schatzman III
STD
0.58876
0.61110
0.39286
0.35512
e=1
MOD
0.97658
1.00200
0.46380
0.46993
e=
1
2
Table 2.1: Convergence rates for the displacement
Method
kλh −λkL2 (0,T )
Oscillations
kEh −EkL∞ (0,T )
kEh −EkL2 (0,T )
STD
-0.13546
Big
-0.61885
-0.60737
MOD
0.48812
Small
0.99486
0.99313
Newmark I
STD
0.01366
Yes
0.50094
0.49943
(β, γ) = ( 12 , 1)
MOD
0.31778
No
0.51230
0.49687
Newmark II
STD
-0.01996
Big
1.05720
1.08500
(β, γ) = ( 21 , 12 )
MOD
0.49128
Small
0.99887
1.00790
STD
0.25559
No
0.50055
0.49963
Scheme
Crank–Nicolson
Backward Euler
MOD
0.30221
No
0.50918
0.49803
Paoli–Schatzman I
STD
0.28976
Yes
1.01620
1.02650
e=0
MOD
0.36928
Yes
0.99331
1.00380
Paoli–Schatzman II
STD
0.30798
Yes
1.33470
1.00160
MOD
0.35044
Yes
1.00550
1.01450
Paoli–Schatzman III
STD
0.04596
Yes
1.02640
0.99142
e=1
MOD
0.34364
Yes
1.01270
1.03150
e=
1
2
Table 2.2: Convergence rates for the Lagrange multiplier and energy
Finally, the convergence rates for different time–stepping methods, analyzed in details above,
are summarized in Tables 2.1 and 2.2 for the reader convenience. These convergence rates were
obtained for both standard (STD) and modified (MOD) mass matrices.
71
Chapter 2. Numerical approximations of a one dimensional elastodynamic contact
problem based on mass redistribution method
One can see in Table 2.1, that the rate of convergence of order one methods (Newmark I
and backward Euler) in Lp (0, T ; H) norms of the displacement (with p = 2, +∞) are limited
to approximatively 21 for both the standard and modified mass matrices. Concerning the other
methods, which are order two methods at least for the linear part of the problem, the rate of
convergence is still approximatively 21 for the standard mass matrix but is improved to approximatively 1 for the modified mass matrix. Note that the Paoli–Schatzman scheme is not a fully
order two method, except for e = 1, due to the discretization of the contact condition.
The advantage is less pronounced in Lp (0, T ; V) norms of the displacement, but it is even
more important in L2 (0, T ) norm of the contact stress (see Table 2.2) since convergence does not
occur for all the methods with the standard mass matrix.
An important remark is that, despite the low regularity of the exact solution, order two
methods perform better than order one methods and should be preferred to approximate elastodynamic problem with impact.
Concerning the efficiency of the mass redistribution method, we can conclude that, even
though it represents an additional approximation compared to the standard approximation, the
convergence rates on the displacement and on the contact force are improved and the method
reduces the potential spurious oscillations. The mass redistribution method appears then to be
a robust method to approximate elastodynamic problem with impact.
2.5.2.1
The Crank–Nicolson method
We focus here on Newmark’s method (2.32) in the particular case where (β, γ) = ( 41 , 21 ).
The analytical solution (u, λ) exhibited in Section 2.5.1 and the approximated solutions
obtained for different time and space steps are represented on Figure 2.7. We observe some spurious oscillations on Figure 2.7 (left) for the solution (Uhn , λn ) to Problem (2.32)
after the contact takes place while these oscillations disappear when the standard mass matrix
is replaced by the modified one, see Figure 2.7 (right).
(Uhn , λn )
It can easily be seen on Figure 2.8 that the scheme with the standard mass matrix is unstable
with a rapidly growing energy while the one with the modified mass matrix is almost conservative
as the space ∆x and time ∆t steps tend to 0. The error curves for (Uhn , λn ) highlighted that
the norms kuh −ukLp (0,T ;H) , with p = 2, +∞, and kλn −λkH converge to 0 when M = M mod as n
tends to +∞, which is unfortunately not the case for the standard mass matrix (see Figure 2.9).
Finally, Figure 2.10 shows that the norm kEh −EkLp (0,T ) , p = 2, +∞, converges when M = M mod
while this norm diverges when the standard mass matrix is considered.
72
2.5. The wave equation with Signorini and Dirichlet boundary conditions
Figure 2.7: Comparison of the analytical solution (u, λ) and the approximated solutions (Uhn , λn )
by using the standard (left) and modified (right) mass matrices in the contact node.
Figure 2.8: Comparison of the energy associated to the analytical solution and the energy associated to the approximated ones for the standard (left) and modified (right) mass matrices.
73
Chapter 2. Numerical approximations of a one dimensional elastodynamic contact
problem based on mass redistribution method
(a) kuh −ukL∞ (0,T ;H)
(b) kuh −ukL2 (0,T ;H)
(c) kuh −ukL∞ (0,T ;V)
(d) kuh − ukL2 (0,T ;V)
(e) kλn −λkL2 (0,T )
Figure 2.9: Comparison of the error curves obtained by using the standard and modified mass
matrices.
74
2.5. The wave equation with Signorini and Dirichlet boundary conditions
(a) kEh −EkL∞ (0,T )
(b) kEh −EkL2 (0,T )
Figure 2.10: Comparison of the error curves of the energy obtained by using the standard and
modified mass matrices.
2.5.2.2
The Newmark method I
We deal now with Newmark’s method (2.32) in the particular case where (β, γ) = ( 21 , 1). We
observe some oscillations on Figure 2.11 (left) for the Lagrange multiplier λn associated to Problem (2.32) after the contact takes place while these oscillations are much less preeminent when
the standard mass matrix is replaced by the modified one, see Figure 2.11 (right). Furthermore,
the present scheme is dissipative and stable.
75
Chapter 2. Numerical approximations of a one dimensional elastodynamic contact
problem based on mass redistribution method
Figure 2.11: Comparison of the analytical solution (u, λ) and the approximated solutions (Uhn , λn )
by using the standard (left) and modified (right) mass matrices in the contact.
Figure 2.12: Comparison of the energy associated to the analytical solution and the energy
associated to the approximated ones for standard (left) and modified (right) mass matrices.
2.5.2.3
The Newmark method II
We consider Newmark’s method (2.32) in the case where (β, γ) = ( 12 , 21 ). Some oscillations on
Figure 2.13 (left) for the Lagrange multiplier λn associated to Problem (2.32) can be observed
after the contact takes place while these oscillations are much less important when the modified mass matrix is considered in (2.32), see Figure 2.13 (right). Note that here the scales for
the Lagrange multiplier associated to standard and modified mass matrices are different, see
Figure 2.13.
76
2.5. The wave equation with Signorini and Dirichlet boundary conditions
Figure 2.13: Comparison of the analytical solution (u, λ) and the approximated solutions (Uhn , λn )
by using the standard (left) and modified mass matrices (right) in the contact.
Figure 2.14: Comparison of the energy associated to the analytical solution and the energy
associated to the approximated ones for the standard (left) and modified (right) mass matrices.
77
Chapter 2. Numerical approximations of a one dimensional elastodynamic contact
problem based on mass redistribution method
2.5.2.4
The backward Euler method
We deal here with backward Euler’s method (i.e. (2.34) with θ = 1). Since the present scheme
is dissipative and stable (∆Ehn < 0, see (2.35)), the approximated solutions (Uhn , λn ) obtained by
using the standard mass as well as modified mass matrices do not oscillate as it can be observed
in the previous numerical simulations.
Figure 2.15: Comparison of the analytical solution (u, λ) and the approximated solutions (Uhn , λn )
by using the standard (left) and modified (right) mass matrices in the contact.
Figure 2.16: Comparison of the energy associated to the analytical solution and the energy
associated to the approximated ones for the standard (left) and modified (right) mass matrices.
78
2.5. The wave equation with Signorini and Dirichlet boundary conditions
2.5.2.5
The Paoli–Schatzman method I
We are interested here in Paoli–Schatzman’s method with (β, e) = ( 14 , 0). Note that the stability
result straightforwardly follows from [124].
Figure 2.17: Comparison of the analytical solution (u, λ) and the approximated solutions (Uhn , λn )
by using the standard (left) and modified (right) mass matrices in the contact.
Figure 2.18: Comparison of the energy associated to the analytical solution and the energy
associated to the approximated ones for the standard (left) and modified (right) mass matrices.
79
Chapter 2. Numerical approximations of a one dimensional elastodynamic contact
problem based on mass redistribution method
2.5.2.6
The Paoli–Schatzman method II
We focus here on Paoli–Schatzman’s method with (β, e) = ( 41 , 12 ). Observe that the stability
result is still an open problem for e 6= 0. Some spurious oscillations can be observed for the
approximated Lagrange multiplier as well as for the approximated energy.
Figure 2.19: Comparison of the analytical solution (u, λ) and the approximated solutions (Uhn , λn )
by using the standard (left) and modified (right) mass matrices in the contact.
Figure 2.20: Comparison of the energy associated to the analytical solution and the energy
associated to the approximated ones for the standard (left) and modified (right) mass matrices.
80
2.5. The wave equation with Signorini and Dirichlet boundary conditions
2.5.2.7
The Paoli–Schatzman method III
We consider now Paoli–Schatzman’s method with (β, e) = ( 14 , 1). Note that this is a second
order scheme.
Figure 2.21: Comparison of the analytical solution (u, λ) and the approximated solutions (Uhn , λn )
by using standard (left) and modified (right) mass matrices in the contact.
Figure 2.22: Comparison of the energy associated to the analytical solution and the energy
associated to the approximated ones for the standard (left) and modified (right) mass matrices.
81
Chapter 2. Numerical approximations of a one dimensional elastodynamic contact
problem based on mass redistribution method
Appendix D
The aim of the present section consists to establish the energy evolutions associated to schemes
(2.32), (2.34) and (2.36). Let us observe that the energy evolution, when the external force
f = 0, is given by
(2.43)
1
1 n+1
n+1
n T
n
−Uh,t
) M (Uh,t
+Uh,t
) + (Uhn+1 −Uhn )T S(Uhn+1 +Uhn ).
∆Ehn = (Uh,t
2
2
We prove now that the energy evolution associated to the Crank–Nicolson method is given
by(2.33). To this aim, we observe that (2.32a) and (2.32b) lead to
(2.44)
(( 1
)
)
)
)
(( 1
n
M (Uhn+1 −Uhn −∆tUh,t
) = −∆t2
−β SUhn +βSUhn+1 − ∆t2
−β λn e0 +βλn+1 e0 ,
2
2
and
(2.45)
n+1
n
M (Uh,t
−Uh,t
) = −∆t((1−γ)SUhn +γSUhn+1 ) − ∆t((1−γ)λn e0 +γλn+1 e0 ).
Then we multiply (2.45) by − ∆t
2 and we subtract this expression to (2.44), we get
(2.46)
n+1
n
M (Uh,t
+Uh,t
)
)
((
)
((
2
γ)
γ)
M (Uhn+1 −Uhn ) + 2∆t β− S(Uhn+1 −Uhn ) + 2∆t β− (λn+1 e0 −λn e0 ) .
=
∆t
2
2
By using (2.46) in (2.43), it follows that
(
1
γ ) n+1
n+1
n T
n T
∆Ehn =
(Uh,t
−Uh,t
) M (Uhn+1 −Uhn ) + ∆t β− (Uh,t
−Uh,t
) S(Uhn+1 −Uhn )
∆t
2
(
1
γ ) n+1
n T n+1
−Uh,t
) λ
e0 −λn e0 ) + (Uhn+1 −Uhn )T S(Uhn+1 +Uhn ).
+ ∆t β− (Uh,t
2
2
Since M is a symmetric matrix, it follows from (2.45) that
(
)
1
∆E = ( − γ)(Uhn+1 − Uhn )T S(Uhn+1 − Uhn ) − (Uhn+1 − Uhn )T (1 − γ)λn e0 + γλn+1 e0
2
γ
γ
n+1
n+1
n T
n T n+1
− Uh,t
) S(Uhn+1 − Uhn ) + ∆t(β − )(Uh,t
− Uh,t
) (λ
e0 − λn e0 ).
∆t(β − )(Uh,t
2
2
By setting the parameters (β, γ) = ( 41 , 12 ) (2.33) holds.
Let us establish the energy evolution associated to the θ–method. We infer from (2.34a) and
(2.34c) that
(2.47)
n+1
n
M (Uh,t
−Uh,t
) = (1−θ)∆t(−SUhn −λn e0 ) + θ∆t(−SUhn+1 −λn+1 e0 ).
Then using (2.47) in (2.43), we find
∆Ehn = −
∆t n+1
n T
(Uh,t +Uh,t
) ((1−θ)SUhn +θSUhn+1 +((1−θ)λn +θλn+1 )e0 )
2
1
+ (Uhn+1 −Uhn )T S(Uhn+1 +Uhn ).
2
82
2.5. The wave equation with Signorini and Dirichlet boundary conditions
Note that from (2.34a)
(1 )
∆t n+1
n+1
n
n
(Uh,t +Uh,t
) = Uhn+1 −Uhn + ∆t −θ (Uh,t
−Uh,t
), it follows that
2
2
(1 )
∆Ehn =
−θ (Uhn+1 −Uhn )T S(Uhn+1 −Uhn ) − (Uhn+1 −Uhn )T (((1−θ)λn +θλn+1 )e0 )
2
(1 )
n+1
n T
+ ∆t −θ (Uh,t
−Uh,t
) ((1−θ)(−SUhn −λn e0 ) + θ(−SUhn+1 −λn+1 e0 )).
2
n+1
n ), which allows
−Uh,t
Clearly we have ∆t((1−θ)(−SUhn −λn e0 ) + θ(−SUhn+1 −λn+1 e0 )) = M (Uh,t
us to conclude that (2.35) holds.
Finally we establish the energy evolution associated to (2.36) in the particular case where
β = 14 . To this aim, let us introduce the energy evolution as follows:
∆Ehn
=
[1
2
T
Uh,t
M Uh,t
]n+ 1
1 T
2
+ Uh SUh
2
n− 12
which leads to
1 n+ 1
1 n+ 1
n− 1
n+ 1
n− 1
n− 1
n+ 1
n− 1
∆Ehn = (Uh,t 2 −Uh,t 2 )T M (Uh,t 2 +Uh,t 2 ) + (Uh 2 −Uh 2 )T S(Uh 2 +Uh 2 ),
2
2
where
n+ 12
Uh
def
=
1 n
n− 1 def 1
(Uh +Uhn+1 ) and Uh 2 = (Uhn +Uhn−1 ).
2
2
Then it comes that
n+ 12
Uh
n− 12
+ Uh
1
1
n− 1
n+ 1
= (Uhn+1 +2Uhn +Uhn−1 ) and Uh 2 − Uh 2 = (Uhn+1 −Uhn−1 ),
2
2
and also
n+ 12
Uhn+1 = Uhn + ∆tUh,t
n− 1
and Uhn = Uhn−1 + ∆tUh,t 2 ,
so we have
n+ 12
Uh,t
n− 12
− Uh,t
=
Uhn+1 −2Uhn +Uhn−1
∆t
n+ 12
and Uh,t
n− 12
+ Uh,t
=
Uhn+1 −Uhn−1
.
∆t
Obviously, we get
( U n+1 −2U n +U n−1 )
( U n+1 +2U n +U n−1 )
1
h
h
h
h
h
∆Ehn = (Uhn+1 −Uhn−1 )T M h
+
S
.
2
∆t2
4
According to (2.36) together with β = 14 , we deduce that
1
1
∆Ehn = (Uhn+1 − Uhn−1 )T (−λn+1 e0 ) = (un+1
− un−1
)(−λn+1 ).
0
2
2 0
n−1
Since un+1
− un−1
= (1+e)un,e
, we find
0
0
0 − u0
(
) 1
1
n+1 n−1
u0 ) = (1+e)λn+1 u0n−1 .
∆Ehn = (1+e) (−λn+1 un,e
0 ) − (−λ
2
2
83
Chapter 2. Numerical approximations of a one dimensional elastodynamic contact
problem based on mass redistribution method
Appendix E
In this section, we are concerned with the convergence curves which are obtained for different
time integration methods. These results have been summarized in the Tables 2.1 and 2.2.
Convergence curves for the Newmark method I
(a) kUhn − ukL∞ (0,T ;L2 (0,L))
(b) kUhn − ukL2 (0,T ;L2 (0,L))
(c) kUhn − ukL∞ (0,T ;H1 (0,L))
(d) kUhn − ukL2 (0,T ;H1 (0,L))
(e) kλn − λkL2 (0,T )
Figure 2.23: Comparison of the convergence curves obtained by using standard and modified
mass matrices.
84
2.5. The wave equation with Signorini and Dirichlet boundary conditions
(a) In kEh − EkL∞ (0,T )
(b) In kEh − EkL2 (0,T )
Figure 2.24: Comparison of the convergence curves of the energy obtained by using standard
and modified mass matrices.
Convergence curves for the Newmark method II
(a) kUhn − ukL∞ (0,T ;L2 (0,L))
(b) kUhn − ukL2 (0,T ;L2 (0,L))
(c) kUhn − ukL∞ (0,T ;H1 (0,L))
85
Chapter 2. Numerical approximations of a one dimensional elastodynamic contact
problem based on mass redistribution method
(d) kUhn − ukL2 (0,T ;H1 (0,L))
(e) kλn − λkL2 (0,T )
Figure 2.25: Comparison of the convergence curves obtained by using standard and modified
mass matrices.
(a) kEh − EkL∞ (0,T )
(b) kEh − EkL2 (0,T )
Figure 2.26: Comparison of the convergence curves of the energy obtained by using standard
and modified mass matrices.
Convergence curves for the backward Euler method
(a) kUhn − ukL∞ (0,T ;L2 (0,L))
(b) kUhn − ukL2 (0,T ;L2 (0,L))
86
2.5. The wave equation with Signorini and Dirichlet boundary conditions
(c) kUhn − ukL∞ (0,T ;H1 (0,L))
(d) kUhn − ukL2 (0,T ;H1 (0,L))
(e) kλn − λkL2 (0,T )
Figure 2.27: Comparison of the convergence curves obtained by using standard and modified
mass matrices.
(a) kEh − EkL∞ (0,T )
(b) kEh − EkL2 (0,T )
Figure 2.28: Comparison of the convergence curves of the energy obtained by using standard
and modified mass matrices.
87
Chapter 2. Numerical approximations of a one dimensional elastodynamic contact
problem based on mass redistribution method
Convergence curves for the Paoli–Schatzman method I
(a) kUhn − ukL∞ (0,T ;L2 (0,L))
(b) kUhn − ukL2 (0,T ;L2 (0,L))
(c) kUhn − ukL∞ (0,T ;H1 (0,L))
(d) kUhn − ukL2 (0,T ;H1 (0,L))
(e) kλn − λkL2 (0,T )
Figure 2.29: Comparison of the convergence curves obtained by using standard and modified
mass matrices.
88
2.5. The wave equation with Signorini and Dirichlet boundary conditions
(a) kEh − EkL∞ (0,T )
(b) kEh − EkL2 (0,T )
Figure 2.30: Comparison of the convergence curves of the energy obtained by using standard
and modified mass matrices.
Convergence curves for the Paoli–Schatzman method II
(a) kUhn − ukL∞ (0,T ;L2 (0,L))
(b) kUhn − ukL2 (0,T ;L2 (0,L))
(c) kUhn − ukL∞ (0,T ;H1 (0,L))
(d) kUhn − ukL2 (0,T ;H1 (0,L))
89
Chapter 2. Numerical approximations of a one dimensional elastodynamic contact
problem based on mass redistribution method
(e) kλn − λkL2 (0,T )
Figure 2.31: Comparison of the convergence curves obtained by using standard and modified
mass matrices.
(a) kEh − EkL∞ (0,T )
(b) kEh − EkL2 (0,T )
Figure 2.32: Comparison of the convergence curves of the energy obtained by using standard
and modified mass matrices.
Convergence curves for the Paoli–Schatzman method III
(a) kUhn − ukL∞ (0,T ;L2 (0,L))
(b) kUhn − ukL2 (0,T ;L2 (0,L))
90
2.5. The wave equation with Signorini and Dirichlet boundary conditions
(c) kUhn − ukL∞ (0,T ;H1 (0,L))
(d) kUhn − ukL2 (0,T ;H1 (0,L))
(e) kλn − λkL2 (0,T )
Figure 2.33: Comparison of the convergence curves obtained by using standard and modified
mass matrices.
(a) kEh − EkL∞ (0,T )
(b) kEh − EkL2 (0,T )
Figure 2.34: Comparison of the convergence curves of the energy obtained by using standard
and modified mass matrices.
91
Chapter 3
Efficiency of a weighted mass
redistribution method for an
elastodynamic problem with unilateral
constraints
Contents
3.1
Introduction
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
93
3.2
Finite element discretization and convergence of the mass redistributions method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
95
3.3
3.1
The wave equation with Signorini and Dirichlet boundary conditions . 101
3.3.1
Analytical solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
3.3.2
Comparisons between different mass redistributions for some time–space discretisations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
3.3.2.1
The Newmark methods . . . . . . . . . . . . . . . . . . . . . . . . 102
3.3.2.2
The backward Euler method . . . . . . . . . . . . . . . . . . . . . 104
3.3.2.3
The Paoli–Schatzman methods . . . . . . . . . . . . . . . . . . . . 105
Introduction
The motion of an elastic bar of length L which is free to move as long as it it does not hit a
material obstacle is studied, see Figure 3.1. The assumptions of small deformations are assumed
and the material of the bar is supposed to be homogeneous. Let u(x, t) be the displacement
at time t ∈ [0, T ], T > 0 of the material point of spatial coordinate x ∈ [0, L]. Let f (x, t)
denotes a density of external forces, depending on time and space. The mathematical problem
is formulated as follows:
(3.1)
utt (x, t) − uxx (x, t) = f (x, t),
(x, t) ∈ (0, L) × (0, T ),
with Cauchy initial data
(3.2)
u(x, 0) = u0 (x) and ut (x, 0) = v 0 (x),
x ∈ (0, L),
Chapter 3. Efficiency of a weighted mass redistribution method for an
elastodynamic problem with unilateral constraints
L
x=0
Figure 3.1: An elastic bar vibrating on impacting obstacle.
and Signorini and Dirichlet boundary conditions at x = 0 and x = L, respectively,
(3.3)
0 ≤ u(0, t) ⊥ ux (0, t) ≤ 0 and u(L, t) = 0,
t ∈ [0, T ].
The orthogonality has the natural meaning: an appropriate duality product between two terms
of relation vanishes.
Let us describe the weak formulation associated to (3.1)–(3.3). For that purpose, it is condef
def
venient to introduce the following notations: V = {u ∈ H1 (0, L) : u(L) = 0}, H = L2 (0, L),
def
def
V = {u ∈ L2 (0, T ; V) : ut ∈ L2 (0, T ; H)} and the convex set K = {u ∈ V : u(0) ≥ 0}. Thus
the weak formulations associated to (3.1)–(3.3) obtained by multiplying (3.1) by v − u and by
def
integrating formally this result over QT = (0, L) × (0, T ) to get
(3.4)


find u ∈ K such that


∫
∫ L



0
0

ut (x, t)(vt (x, t)−ut (x, t)) dx dt+
v (x)(v(x, 0)−u (x)) dx −
−
QT ∫
∫ 0


ux (x, t)(vx (x, t)−ux (x, t)) dx dt ≥
f (x, t)(v(x, t)−u(x, t)) dx dt



QT
QT



for all v ∈ K for which there exists ζ > 0 with v = u for t ≥ T − ζ.
Existence and uniqueness results are obtained for a similar situation of a vibrating string with
concave obstacle in one dimensional space in [40] and also for a wave equation with unilateral
constraint at the boundary in a half–space of RN in [44]. An existence result for a wave equation
in a C2 –regular bounded domain constrained by an obstacle at the boundary in RN for N ≥ 2
is proven in [45]. The reader is also referred to [88].
The paper is organized as follows. In Section 3.2, a space semi–discretization based on mass
redistribution method is presented. This mass redistribution method consists to transfer the
mass of the contact node on the other nodes while the inertia vanishes at the contact node. The
error estimate in time as well as the convergence result are established. A benchmark problem
is introduced in Section 3.3 and its analytical solution is exhibited. Then numerical experiments
for some space–time discretizations like the Crank–Nicolson or the backward Euler methods are
reported. These numerical experiments highlight that the choice of the nodes where the mass is
transferred plays a crucial role to get a better approximate solution.
94
3.2. Finite element discretization and convergence of the mass redistributions
method
3.2
Finite element discretization and convergence of the mass
redistributions method
This section is devoted to semi–discrete problems in space associated to (3.4) by using the mass
redistributions method, see [84, 88]. More precisely, the mass redistributions method consists
in transferring the mass of the contact node on the other nodes implying that the node at the
contact boundary evolves in a quasi–static way. To this aim, we choose a set of parameters
def L
h = m
(mesh size) having in mind the limit m tends to +∞, where m is an integer, and let
def
Vh = {vh ∈ C0 ([0, L]) : vh |[ih,(i+1)h] ∈ P1 , i = 0, . . . , m − 1, vh (L) = 0} where P1 is the space of
polynomials of degree less than or equal to 1. A classical basis of Vh is given by the sequence of
shape functions ϕi ∈ Vh for i = 0, 1, . . . , J, defined by
{
1 − |x−ih|
if x ∈ [amax(i−1,0) , amin(i+1,m) ],
def
h
ϕi (x) =
0
otherwise,
where J = m − 1. Clearly, ϕi (jh) = δij , j = 0, 1, . . . , m, where δ denotes Kronecker’s symbol.
Thus the solution u belonging to V to the weak formulation (3.4) is approximated by
uh (x, t) =
J
∑
uj (t)ϕj (x).
j=0
Consequently, ui = uh (ai ), i = 0, 1, . . . , n − 1 and the weak formulation (3.4) is approximated by


find uh : [0, T ] → Vh and λh : [0, T ] → R such that for all vh ∈ Vh


∫ L
∫ L
∫ L




f vh dx a.e. t ∈ [0, T ],
uh,x vh,x dx = −λh vh (0) +
uh,tt vh dx +
(Puh )
0
0
0


0 ≤ uh (0, ·) ⊥ λh ≤ 0 a.e. t ∈ [0, T ],



u (·, 0) = u0 and u (·, 0) = v 0 ,
h
h
h,t
h
where u0h and vh0 belong to Vh and λ is the Lagrange multiplier being here the contact force.
The corresponding algebraic formulation is given by


find Uh : [0, T ] → Rn and λh : [0, T ] → R such that



M U + SU = −λe + F a.e. t ∈ [0, T ],
0
h,tt
h
(PUh λh )

0 ≤ u0 ⊥ λh ≤ 0 a.e. t ∈ [0, T ],



Uh (0) = U 0 and Uh,t (0) = V 0 ,
def
where M and S denote the mass
and stiffness∫matrices, respectively, and Uh = (u0 , . . . , un−1 )T ,
def
def ∫ L
L
T
e0 = (1, 0, . . . , 0) and F = ( 0 f ϕ0 dx, . . . , 0 f ϕn−1 dx)T with (·)T being the transpose of a
tensor.
An alternative approach to the standard discretization presented above is to consider the
mass redistributions method which consists to replace the mass matrix M in (PUh λh ) by a
def ∫ L
modified mass matrix M mod defined by Mijmod = 0 ϕi ϕj wh,ij dx such that wh,0i = wh,i0 for all
95
Chapter 3. Efficiency of a weighted mass redistribution method for an
elastodynamic problem with unilateral constraints
i = 0, . . . , n − 1 and wh,ij = 1 + o(h) for all i, j = 1, . . . , n − 1. Clearly this modified mass matrix
reads
(
)
0 0
mod def
=
.
M
¯
0 M
This leads to an algebraic formulation of the semi–discrete approximation with mass redistribution method given by


find Uh : [0, T ] → Rn and λ : [0, T ] → R such that




¯¯
¯¯
¯


 M Uh,tt + S Uh = −Φ0 u0 + F a.e. t ∈ [0, T ],
¯
(Pmod
S00 u0 + ΦT
a.e. t ∈ [0, T ],
U h λh )
0 Uh = −λh + F0




0 ≤ u0 ⊥ λh ≤ 0 a.e. t ∈ [0, T ],



 U (0) = U 0 and U (0) = V 0 .
h
h,t
def
def ∫
def
¯ def
Here U
= (u1 , . . . , un−1 )T , S¯ij = Si+1,j+1 with Φ0 = Ω ϕ0i+1 ϕ00 dx, i = 0, . . . , n − 2, F¯ =
( −λh −ΦT U¯h )
∫L
∫L
0
=
( 0 f ϕ1 dx, . . . , 0 f ϕn−1 dx)T . Since Φ0 = (S10 , 0, . . . , 0)T , it comes that u0 =
S00
( −λ −S10 u1 )
h
. If we assume that S10 u1 ≥ 0 then the compatibility condition implies that u0 = 0
S00
(
)
¯h )− otherwise λ= 0. This gives that u0 = −S10 u1 + , and then we may conclude
and λh = (ΦT U
0
S00
that (Pmod
Uh λh ) is equivalent to the following second order Lipschitz continuous ordinary differential
equation:

¯h : [0, T ] → Rn−1 such that
find U




( −S u )+
10 1
¯
¯
¯
¯
(Pmod
)
M
U
+
S
U
=
−Φ
+ F¯ a.e. t ∈ [0, T ],
0
¯h
h,tt
h
U

S
00



Uh (0) = U 0 and Uh,t (0) = V 0 .
def
¯h , U
¯h,t )T and G(U¯h ) def
It is convenient to introduce the following notations: U¯h = (U
=
def
−1
−1
−1
T
+
T
mod
¯
¯
¯
¯
¯
¯
¯
(Uh,t , −M S Uh + M DU¯h + M F ) with DU¯h = ((u1 ) /h, 0, . . . , 0) . Therefore (PU¯ )
h
can be rewritten as follows:

n−1
¯
× Rn−1 such that

 find Uh : [0, T ] → R
U¯h,t = G(U¯h ) a.e. t ∈ [0, T ],
(Pmod
U¯h )

¯
Uh (0) = (U 0 , V 0 )T .
Observe that G(U¯h ) is Lipschitz continuous. Indeed, let U¯h1 and U¯h2 be two solutions to Problem
), then we have
(Pmod
U¯
h
¯ 2 )kU¯1 − U¯2 k2H ,
kG(U¯h1 ) − G(U¯h2 )k2H ≤ max(1, 1/h2 , |||S|||
h
h
where ||| · ||| denotes the induced matrix norm. Existence and uniqueness results for the problem
) follow from the Lipschitz continuity of G(U¯h ), for further details the reader is referred
(Pmod
U¯h
to [99]. In particular, we have U¯h ∈ C1 ([0, T ]).
T
Lemma 11. Let ∆t be the time step and N = ∆t
∈ N for given T > 0. Then the truncation
error for the Crank–Nicolson method to solve the semi–discrete problem (Pmod
¯ ) is of order ∆t.
U
h
96
3.2. Finite element discretization and convergence of the mass redistributions
method
Proof. Let the time discretization [0, T ] be divided by n + 1 discrete time–points such that
n are assumed
0 = t0 < t1 < . . . < tn = T . Furthermore the discrete quantities U¯hn and U¯h,t
to be given by algorithmic approximations of the displacement U¯h (tn ) and the velocity U¯h,t (tn ),
respectively. Notice that the Crank–Nicolson method is based on the trapezoidal rule in time.
Thus the straightforward application of the Crank–Nicolson method to Problem (Pmod
) is a
U¯h
combination of the forward Euler method at n and the backward Euler method at n + 1, we
obtain
U¯hn+1 − U¯hn
1
= (G(U¯hn+1 ) − G(U¯hn )).
∆t
2
(3.5)
) that
Since U¯ ∈ C1 ([0, T ]), we may deduce from (Pmod
U¯
1
U¯h (tn+1 ) − U¯h (tn )
= (G(U¯h (tn+1 )) − G(U¯h (tn ))) + εn (∆t),
∆t
2
(3.6)
where εn is the local truncation error such that
kεn (∆t)kH ≤
(3.7)
CF 0 ∆t
2
with
def
CF 0 =
¯ H.
sup kF 0 (U)k
[t0 ,t0 +T ]
def
Let us define the global truncation error e¯n = U¯hn − U¯h (tn ). Then we substract (3.6) from (3.5)
to get
e¯n+1 = e¯n +
)
∆t (
(G(U¯hn+1 ) − G(U¯h (tn+1 ))) − (G(U¯hn ) − G(U¯h (tn ))) − εn (∆t2 ).
2
By using the Euclidean inner product (·, ·), we get
)
∆t ( n+1
e¯ , G(U¯hn+1 ) − G(U¯h (tn+1 ))
2
) (
)
∆t ( n+1
−
e¯ , G(U¯hn ) − G(U¯h (tn )) − e¯n+1 , εn (∆t2 ) .
2
(¯
en+1 , e¯n+1 ) =(¯
en+1 , e¯n ) +
(3.8)
¯ where Id denotes the identity matrix. It comes from
Let us define CSh¯ = max(1, 1/h)|||Id − S|||
(3.8) that
(
(
CSh¯ ∆t ) n+1
CSh¯ ∆t ) n
1−
k¯
e
kH ≤ 1 +
k¯
e kH + kεn (∆t2 )kH .
2
2
def
According to (3.7), for ∆t sufficiently small, we find
k¯
e
n+1
(3.9)
where CSh,∆t
= 1−
¯
def
k¯
e
n+1
CSh¯ ∆t
2 .
(
)
CSh¯ ∆t
CF 0 ∆t2
kH ≤ 1 + h,∆t k¯
en k H +
CS¯
2CSh,∆t
¯
We may infer from (3.9) that
)(
)
n (
(
CSh¯ ∆t )n+1 0
CSh¯ ∆t i ∆t2
CF 0 ∑
k¯
e kH +
,
kH ≤ 1 + h,∆t
1 + h,∆t
2
CS¯
CS¯
CSh,∆t
¯
i=0
97
Chapter 3. Efficiency of a weighted mass redistribution method for an
elastodynamic problem with unilateral constraints
which implies that
(
)
( 1 − 1 + CSh¯ ∆t n+1 )(
)
)
(
CSh¯ ∆t n+1 0
CSh,∆t
CF 0
∆t2
¯
n+1
k¯
e kH +
.
k¯
e
kH ≤ 1 + h,∆t
(
)
C h¯ ∆t
2
CS¯
CSh,∆t
¯
1 − 1 + Sh,∆t
CS¯
Then the Taylor expansion leads to
(
k¯
e
n+1
kH ≤ exp
)
(
( h ))
CS¯ T
CF 0 ∆t
0
k¯
e kH +
.
1 − exp
h,∆t
h
2CS
C¯
C h,∆t
¯
CSh¯ T
S
It follows that
k¯
e
n+1
S
( h )(
)
CS¯ T
CF 0 ∆t
0
kH ≤ exp
k¯
e kH +
.
2CSh¯
CSh,∆t
¯
¯ Clearly, we have C h¯ ≥ max(1, 1/h)ρ(Id+
Let us denote the spectral radius of Id+ S¯ by ρ(Id+ S).
S
¯ ≥ max(1, 1/h) and it comes that
S)
k¯
e
n+1
)
( h )(
CS¯ T
CF 0 ∆t
0
k¯
e kH +
.
kH ≤ exp
max(1, 1/h)
CSh,∆t
¯
which concludes the proof.
On the other, we may observe that (Pmod
U λ ) is equivalent to


find uh : [0, T ] → Vh such that for all vh ∈ K ∩ Vh


∫ T∫ L





uh,tt (x, t)(vh (x, t)−uh (x, t))wh (x) dx dt
mod
∫
0∫ h
(Puh )


f (x, t)(vh (x, t)−uh (x, t)) dx dt,
uh,x (x, t)(vh,x (x, t)−uh,x (x, t)) dx dt ≥
+



QT
QT



uh (x, 0) = u0h (x) and uh,t (x, 0) = vh0 (x).
We assume that the initial data u0h and uh,t satisfy
(
)
(3.10)
lim ku0h −u0 kV + kvh0 −v 0 kH = 0.
h→0
The convergence of the solution uh of (Pmod
uh ) to the solution of (3.4) is proved below. To this aim,
the same techniques detailed in the proof of theorem 4.3 in [88] are used. The main difference
comes from the weight functions that are introduced in the present work. Indeed the mass of
the contact node is redistributed on the other nodes by using these weight functions while this
mass vanishes in [88]. The reader is also referred to [89]. Furthermore, we may observe that for
all τ ∈ [0, T ], the following energy relation
∫
∫
L
2
(3.11)
2
0
(|ut (x, τ )| +|ux (x, τ )| ) dx =
0
∫
L
2
(|v (x)|
0
+|u0x (x)|2 ) dx
+
f (x, t)ut (x, t) dx dt
Qτ
holds.
98
3.2. Finite element discretization and convergence of the mass redistributions
method
Theorem 11. Assume that (3.10) holds. Then, the solution uh of (Pmod
) converges in the strong
uh
topology of V to the unique solution of (3.4) as h tends to 0.
Proof. We observe that the energy relation (3.11) and the Gronwall lemma yields that there
exists C > 0 independent of τ such that
sup (kuh (·, τ )kV +kuh,t (·, τ )kH ) ≤ C.
t∈[0,T ]
Then we may extract a subsequence, still denoted by uh , such that
(3.12a)
(3.12b)
L∞ (0, T ; V)
in
uh * u
in
uh,t * ut
weak
∗,
L (0, T ; H) weak
∗.
∞
Let us define( H∞ = {u ∈ L∞ (0, T ;)V) : ut ∈ L∞ (0, T ; H)} endowed with the norm kukH∞ =
ess supt∈[0,T ] ku(·, t)kV + kut (·, t)kH . Then, we may deduce from (3.12) that
def
(3.13)
def
uh * u
in
H∞
weak
∗.
1
Notice that for all α < 12 , we have H∞ ,→ C0, 2 (QT ) ,→,→ C0,α (QT ) hold (see [89]), where ,→
and ,→,→ denote the continuous and compact embeddings, respectively. Finally we find
uh → u
(3.14)
in
C0,α (QT )
for all α < 21 . Furthermore, uh and u belong to K. Our aim consists to establish that the limit
u satisfies (3.4). However the elements of K are not smooth enough in time, then they should
be approximated before being projected onto Vh . Indeed this projection violates the constraint
at x = 0, and therefore, the elements of K need another approximation in order to satisfy the
constraint strictly. Assume that v ∈ K such that v = u for t ≥ T − ε and for η ≤ 4ε , let us define
∫ t+η

u(x, t) + 1
(v(x, s)−u(x, s)) ds + k(η)(L−x)ψ(t)
def
η t
vη (x, t) =

u(x, t)
if
t ≤ T − η,
if
t ≥ T − η,
where ψ is a smooth and positive function defined by ψ = 1 on [0, T − 2ε ] and ψ = 0 on [T − 4ε , T ].
We precise now how the parameter k(η) to ensure that vη ∈ K ∩ L∞ (0, T ; V) is chosen. Since
1
u ∈ C0, 2 (QT ), it follows that there exists Ck > 0, such that
∫
1 ∫ t+η
1 t+η
u(0, s) ds ≤
|u(0, t)−u(0, s)| ds
u(0, t) −
η t
η t
∫
Ck kukH∞ η √
2
√
≤
s ds = Ck kukH∞ η.
η
3
0
Then for all t ≤ T − 2ε , we obtain
1
vη (0, t) ≥
η
∫
t
t+η
2
√
v(0, s) ds − Ck kukH∞ η + k(η)Lψ(t).
3
99
Chapter 3. Efficiency of a weighted mass redistribution method for an
elastodynamic problem with unilateral constraints
Observe that the choice k(η) =
√
5
3L Ck kukH∞ η
ensures that for all t ≤ T − 2ε , we get
√
vη (0, t) ≥ Ck kukH∞ η.
Let ProjVh be the orthogonal projection onto Vh with respect to the scalar product in H. Since
the sequence ProjVh satisfies kProjVh z−zkV → 0 when h → 0 for all z ∈ V, the Sobolev injections
imply that there exists a sequence αh converging to zero as h tends to zero such that for all z
belonging to V, we have
kProjVh z−zkC0 ≤ αh kzkV .
The next step consists in choosing an adequate test function. Let us define
vh (·, t) = uh (·, t) + ProjVh (vη −u)(·, t)
def
(3.15)
for all t belonging to [0, T ]. Clearly, vh ∈ K∩Vh , for all t, and for all h small enough. Introducing
(3.15) into (Pmod
uh ) and according to the integration by parts, it comes that
∫ L
∫
−
uh,t (x, 0)ProjVh (vη −u)(x, 0)wh (x) dx −
uh,t (x, t)ProjVh (vη,t −ut )(x, t)wh (x) dx dt
0
QT
∫
∫
f (x, t)ProjVh (vη −u)(x, t) dx dt.
uh,x (x, t)(ProjVh (vη −u))x (x, t) dx dt ≥
+
QT
QT
We already know that
ProjVh (vη,t −ut ) → vη,t −ut
(3.16a)
in L2 (0, T ; H),
ProjVh (vη −u) → vη −u in L2 (0, T ; V).
(3.16b)
According to (3.12)–(3.16), we are able to pass to the limit as h tends to 0, we find
∫
∫ L
0
ut (x, t)(vη,t −ut )(x, t) dx dt
v (vη −u)(x, 0) dx −
−
0
QT
∫
∫
+
ux (x, t)(vη −u))x (x, t) dx dt ≥
f (x, t)(vη −u)(x, t) dx dt.
QT
QT
Since it is quite a routine to pass to the limit with respect to η to get the variational formulation
(3.4), the verification is left to the reader. It remains to prove that uh converges strongly in V.
Notice that (Pmod
uh ) satisfies the following energy relation
∫ L
(|uh,t (x, τ )|2 wh (x)+|uh,x (x, τ )|2 ) dx
0
∫ L
∫
0
2
0
2
=
(|v (x)| wh (x)+|uh,x (x)| ) dx +
f (x, t)uh,t (x, t) dx dt.
0
Qτ
According to (3.10) and to the convergence results established above, we get
∫ L
lim
(|uh,t (x, τ )|2 wh (x)+|uh,x (x, τ )|2 ) dx
h→0 0
(3.17)
∫ L
∫
0
2
0
2
=
(|v (x)| +|uh,x (x)| ) dx +
f (x, t)ut (x, t) dx dt.
0
Qτ
100
3.3. The wave equation with Signorini and Dirichlet boundary conditions
Therefore (3.11) and (3.17) enable us to infer that
∫ L
∫
lim
(|uh,t (x, τ )|2 wh (x)+|uh,x (x, τ )|2 ) dx =
h→0 0
L
(|ut (x, τ )|2 +|ux (x, τ )|2 ) dx
0
Since uh converges weakly to u in H∞ , kuh kH∞ converges to kukH∞ and H∞ ,→ H2 , we may
conclude that uh converges strongly to u in H2 .
3.3
The wave equation with Signorini and Dirichlet boundary
conditions
We consider a bar of length L = 1 clamped at one end and compressed at t = 0. The bar
elongates under the gravity effect; as soon as it reaches a rigid obstacle at time t1 then it stays
in contact during the time t2 − t1 and it takes off at time t2 , see Figure 3.2. This problem
can be described mathematically by (3.1)–(3.3) with the density of external forces f (x, t) = 0.
We first describe how an analytical piecewise affine and periodic solution to our problem can
be obtained by using the characteristics method, the reader is referred to [111] for a detailed
explanation. Then approximate solutions for some time–space discretizations and for several
mass redistributions are exhibited and their efficiency is discussed.
3.3.1
Analytical solution
The domains considered here are defined by (0, L) × (ti , ti+1 ), i = 0, 1, 2, corresponding to the
phases before, during and after the impact, respectively. Each of them are divided into four
regions as it is represented on Figure 3.2. We choose below ti = i, i = 0, . . . , 3.
x
before
contact
L=1
during
contact
II
I
II
III
IV
after
contact
I
II
III
I
IV
III
IV
t
0
t0 = 0
u = 1−x
2
v=0
t1 = 1
u=0
v = −1/2
t2 = 2
u=0
v = 1/2
t3 = 3
u = 1−x
2
v=0
Figure 3.2: The regions allowing to determine the value of u.
The domain (0, 1) × (0, 1) corresponding to the phase before the impact is split into four regions
according to the characteristics lines x + t and x − t. Therefore
{
1−x
2 in the regions I, III,
u(x, t) = 1−t
(3.18)
2 in the regions II, IV.
101
Chapter 3. Efficiency of a weighted mass redistribution method for an
elastodynamic problem with unilateral constraints
The domain (1, 2) × (0, 1) corresponding to the phase during the impact is also divided into four
regions. Then the solution (3.18) evaluated in the region IV allows us to infer that ut (·, 1) = − 21
and we conclude that

t−t1
x


 2 in the region I, − 2 in the region II,
(3.19)
u(x, t) = x−1
2 in the region III,


 t−t1 −1 in the region IV.
2
The domain (2, 3) × (0, 1) corresponding to the phase after the impact is split into four regions.
By using (3.19), we get u(·, 2) = 0 and ut (·, 2) = 12 which leads to
{
t−t2
2 in the regions I, II,
u(x, t) = 1−x
2 in the regions III, IV.
Since u(·, 3) = u0 and ut (·, 3) = v 0 , the solution u(x, t) is periodic of period 3. Finally, note that
λ = ux (0, ·).
3.3.2
Comparisons between different mass redistributions for some time–
space discretisations
The time discretization is introduced in this section. To this aim, we divide the time interval
n ,
[0, T ] by n + 1 discrete time–points such that 0 = t0 < t1 < . . . < tn = T and let Uhn , Uh,t
n and λn the approximations of the displacement U (t ), the velocity U (t ), the acceleraUh,tt
h n
h,t n
tion Uh,tt (tn ) and the Lagrange multiplier λ(tn ), respectively. We deal with some approximate
solutions to Problem (3.1)–(3.3) obtained by using several time–stepping methods like the Newmark, backward Euler and Paoli–Schatzman methods. For each of these time–stepping methods,
approximate solutions (Uhn , λn ) are exhibited for several mass redistributions and they are compared to the analytical solution (u, λ) introduced in Section 3.3.1. In the numerical experiments
presented below, Mod1, Mod2 and Mod3 correspond to the mass of the contact node that is
not redistributed, that is redistributed on all the other nodes and that is redistributed on the
node preceding the contact node, respectively. The efficiency of the mass redistribution method
depends on the position of the nodes where the mass is redistributed. Indeed, the numerical
experiments highlight that the closer from the contact node the mass is transferred the better
the approximate solutions. Then, it is not surprising that the best approximate solution can
be expected and indeed it is obtained in the case where all the mass of the contact node is
transferred on the node preceding the contact node, see Figures 3.3, 3.5 and 3.7. Note that the
numerical simulations presented below were performed by employing the finite element library
Getfem++ (see [87]).
3.3.2.1
The Newmark methods
The Taylor expansions of displacements and velocities neglecting terms of higher order are the
underlying concept of the family of Newmark methods, see [74]. These methods are unconditionally stable for linear elastodynamic problem for γ ≥ 12 and β ≥ 14 ( 12 +γ)2 , see [120, 121], but
they are also the most popular time–stepping schemes used to solve contact problems.
102
3.3. The wave equation with Signorini and Dirichlet boundary conditions
Exact solution
Mod 1
Mod 2
Mod 3
0.6
0.5
Displacement in the contact point
Displacement in the contact point
0.7
0.4
0.3
0.2
0.1
0
Exact solution
Mod 1
Mod 2
Mod 3
0.4
0.3
0.2
0.1
0
−0.1
−0.2
0
1
2
3
Time
4
5
−0.1
0
6
0.2
0.4
0.6
Time
0.8
1
1.2
0.6
Exact solution
Mod 1
Mod 2
Mod 3
0.4
Lagrange multiplier (λ)
Lagrange multiplier (λ)
0.5
Exact solution
Mod 1
Mod 2
Mod 3
0
−0.5
−1
0.2
0
−0.2
−0.4
−0.6
−0.8
−1.5
0
1
2
3
Time
4
5
−1
0.8
6
0.85
0.9
0.95
Time
1
1.05
1.1
Figure 3.3: Comparison of the analytical solution (u, λ) and the approximate solutions (Uhn , λn )
by modified mass matrices in the contact node.
0.2
Total Energy
0.15
Exact solution
Mod 1
Mod 2
Mod 3
0.1
0.05
0
0
1
2
3
Time
4
5
6
Figure 3.4: Comparison of the energy associated to the analytical solution and the energy associated to approximate solutions for modified mass matrices.
103
Chapter 3. Efficiency of a weighted mass redistribution method for an
elastodynamic problem with unilateral constraints
The discrete evolution for the contact problem (3.1)–(3.3) is described by the following finite
difference equations:


find Uhn+1 : [0, T ] → Rn and λn : [0, T ] → R such that:



)
(

n+1
n + ∆tU n + 1 −β ∆t2 U n + β∆t2 U n+1 ,

U
=
U

h,tt
h
h,t
h,tt
h
2

(3.20)
n+1
n + (1−γ)∆tU n + γ∆tU n+1 ,
= Uh,t
Uh,t
h,tt
h,tt


n+1
n+1

n+1 e + F n+1 ,

M
U
+
SU
=
−λ
0

h,tt
h


0 ≤ un+1 ⊥ λn+1 ≤ 0,
0
where ∆t is a given times step and (β, γ) are the algorithmic parameters, see [120, 112]. Note
0 and λ0 are given and U 0
that Uh0 , Uh,t
h,tt is evaluated by using the third identity in (3.20). We
are particularly interested in the case where (β, γ) = ( 41 , 12 ). This method is called the Crank–
Nicolson method, it is second–order consistent and unconditionally stable in the unconstrained
case. However the situation is quite different in the case of contact constraints, indeed the order
of accuracy is degraded, for further details, the reader is referred to [120, 121, 122]. The analytical
solution (u, λ) exhibited in Section 3.3.1 and the approximate solutions (Uhn , λn ) obtained for
different mass redistributions are represented on Figure 3.3. The approximate solution obtained
for the redistribution of the contact node mass on the node preceding the contact (see Mod 3 on
Figure 3.3) gives much better accuracy than the mass redistribution on all the nodes preceding
the contact node (see Mod 2 on Figure 3.3) and no mass redistribution. This highlighted that
the choice for the mass redistribution plays a crucial role.
3.3.2.2
The backward Euler method
We are concerned here with backward Euler’s method which for the contact problem (3.1)–(3.3)
is described by the following finite difference equations:


find Uhn+1 : [0, T ] → Rn and λn : [0, T ] → R such that:




n+1
n+1
n


Uh = Uh + ∆tUh,t ,
n+1
n + ∆tU n+1 ,
(3.21)
Uh,t
= Uh,t
h,tt


n+1
n+1


M Uh,tt + SUh = −λn+1 e0 + F n+1 ,



0 ≤ un+1 ⊥ λn+1 ≤ 0.
0
0 and λ0 are given and U 0
Note that Uh0 , Uh,t
h,tt is evaluated by using the third equality in (3.21).
104
3.3. The wave equation with Signorini and Dirichlet boundary conditions
0.6
0.5
0.5
Exact solution
Mod 1
Mod 2
Mod 3
Displacement in the contact point
Displacement in the contact point
0.7
0.4
0.3
0.2
0.1
0
−0.1
−0.2
0
1
2
3
Time
4
5
Exact solution
Mod 1
Mod 2
Mod 3
0.4
0.3
0.2
0.1
0
−0.1
0
6
0.2
0.4
0.6
Time
0.8
1
1.2
0.95
Time
1
1.05
1.1
0.6
Exact solution
Mod 1
Mod 2
Mod 3
0.4
Lagrange multiplier (λ)
Lagrange multiplier (λ)
0.5
Exact solution
Mod 1
Mod 2
Mod 3
0
−0.5
−1
0.2
0
−0.2
−0.4
−0.6
−0.8
−1.5
0
1
2
3
Time
4
5
−1
0.8
6
0.85
0.9
Figure 3.5: Comparison of the analytical solution (u, λ) and the approximate solutions (Uhn , λn )
by modified mass matrices in the contact node.
0.2
0.18
0.16
Exact solution
Mod 1
Mod 2
Mod 3
Total Energy
0.14
0.12
0.1
0.08
0.06
0.04
0.02
0
0
1
2
3
Time
4
5
6
Figure 3.6: Comparison of the energy associated to the analytical solution and the energy associated to approximate solutions for modified mass matrices.
3.3.2.3
The Paoli–Schatzman methods
We focus on the so–called Paoli–Schatzman method that consists to fix the contact constraint
at an intermediate time step. Indeed the method proposed below is a slight modification of
105
Chapter 3. Efficiency of a weighted mass redistribution method for an
elastodynamic problem with unilateral constraints
Paoli–Schatzman method (see [85, 123]) which takes into account the kernel of the modified
mass matrix. A simple application of Paoli–Schatzman method based on Newmark scheme to
Problem (PU¯h ,λh with γ = 21 leads to


find Uhn+1 : [0, T ] → Rn and λn : [0, T ] → R such that:





M (Uhn+1 −2Uhn +Uhn−1 )


+ S(βUhn+1 +(1−2β)Uhn +βUhn−1 ) = −λn e0 for all n ≥ 2,
2
∆tn+1
(3.22)
u0 + eun−1

n,e
0


0
≤
u
=
⊥ λn ≤ 0,

0

1
+
e


U and U given.
0
1
Here e belongs to [0, 1] and is aimed to be interpreted as a restitution coefficient. Note that Uh0
and Uh1 are given data and Uh1 can be evaluated by a one step scheme. We may observe that
taking M = M mod in (3.22), we are not able to resolve the problem on the kernel of M mod . That
is the reason why, SUhn−1 as well as SUhn are projected on the orthogonal of the kernel of M .
We are interested here in Paoli–Schatzman’s method with (β, e) = ( 41 , 0). Note that the stability
result straightforwardly follows from [124].
0.6
0.5
0.5
Exact solution
Mod 1
Mod 2
Mod 3
Displacement in the contact point
Displacement in the contact point
0.7
0.4
0.3
0.2
0.1
0
−0.1
−0.2
0
1
2
3
Time
4
5
0.4
0.3
0.2
0.1
0
−0.1
0
6
Exact solution
Mod 1
Mod 2
Mod 3
0.2
0.4
0.6
Time
0.8
1
1.2
0.95
Time
1
1.05
1.1
0.6
0.4
Lagrange multiplier (λ)
Lagrange multiplier (λ)
0.5
Exact solution
Mod 1
Mod 2
Mod 3
0
−0.5
−1
0.2
Exact solution
Mod 1
Mod 2
Mod 3
0
−0.2
−0.4
−0.6
−0.8
−1.5
0
1
2
3
Time
4
5
−1
0.8
6
0.85
0.9
Figure 3.7: Comparison of the analytical solution (u, λ) and the approximate solutions (Uhn , λn )
by modified mass matrices in the contact node.
106
3.3. The wave equation with Signorini and Dirichlet boundary conditions
0.2
Total Energy
Exact solution
Mod 1
Mod 2
Mod 3
0.15
0.1
0.05
0
1
2
3
Time
4
5
6
Figure 3.8: Comparison of the energy associated to the analytical solution and the energy associated to approximate solutions for modified mass matrices.
107
Chapter 4
Numerical study of convergence of the
mass redistribution method for
elastodynamic contact problems
This chapter is published in the proceedings of WCCM–ECCM–ECFD 2014 [125]
Contents
4.1
Introduction
4.2
The finite element approximation of problem
4.3
4.4
4.1
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
. . . . . . . . . . . . . . . 111
4.2.1
The semi–discretization in space . . . . . . . . . . . . . . . . . . . . . . . . 111
4.2.2
The time integration methods . . . . . . . . . . . . . . . . . . . . . . . . . . 112
Numerical experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
4.3.1
The Crank–Nicolson scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
4.3.2
The backward Euler method . . . . . . . . . . . . . . . . . . . . . . . . . . 120
Some complements
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
4.4.1
The Newmark method I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
4.4.2
The Newmark method II . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
4.4.3
The convergence rates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
Introduction
This note aims to give some numerical results for the evolution of an elastic material being
subjected to unilateral boundary conditions by employing the mass redistribution method introduced in [84]. The contact phenomena is modelled by using the so–called Signorini’s boundary
conditions in displacement, which are based on a linearization of the physically meaningful non
penetrability of masses. A considerable engineering and mathematical literature is devoted to
contact problems, however a few existence of solutions results have been established, the reader
is referred to [44, 45]. Therefore a challenging task consists to elaborate efficient numerical
methods able to approximate this problem. The present work adopts the approach consisting to
remove the mass at the contact nodes which prevents the oscillations at the contact boundaries
as already observed for the one dimensional elastodynamic contact problems in [111].
We consider an elastic bar vibrating vertically such that one end of this bar is free to move,
as long as it does not hit a material obstacle, while the other end is clamped . The obstacle
constrains the displacement of the extremity to be greater than or equal to 0. We have supposed
Chapter 4. Numerical study of convergence of the mass redistribution method for
elastodynamic contact problems
that the material of the bar is a homogeneous and isotropic. Let us define Ωd ⊂ Rd , d = 2, 3,
Dir
Neu
Sig
def
the reference configuration of the bar and its boundary by ∂Ωd = Γd ∪ Γd ∪ Γd where ΓDir
d ,
Sig
ΓNeu
and
Γ
denote
the
Dirichlet,
Neumann
and
unilateral
contact
boundaries,
respectively.
d
d
def
def
In our case, we assume that Ω2 = [0, L1 ] × [0, L2 ] and Ω3 = [0, L1 ] × [0, L2 ] × [0, L3 ] together
with Li > 0, i = 1, 2, 3. We denote by u(x, t) the displacement field at time t ∈ [0, T ], T > 0,
of the material point of spatial coordinate x ∈ Ω ⊂ Rd . Furthermore, let f (x, t) be the density
of external forces depending on time and space. The mathematical problem is formulated as
follows:
(4.1)
ρutt − div σ(u) = f
in Ωd × (0, T ),
where ρ > 0 is the mass density, (·)t = ∂(·)
∂t and σ(u) is the stress tensor with σij (u) = Aijkl εkl (u)
def 1 ∂uk
∂ul
where Aijkl and εkl (u) = 2 ( ∂xl + ∂xk ) are the stress and strain tensors, respectively. Here the
summation convention on repeated indices is used. Besides assuming isotropy of the material,
def
the Hooke tensor is defined via Lamé constants λ and µ as Aijkl = λδij δkl + 2µδkl δjk where δij
denotes the Kronecker symbol. The Cauchy initial data are
def
(4.2)
u(·, 0) = u0
def
and ut (·, 0) = v 0
on Ωd ,
and the boundary conditions are
(4.3a)
Neu
u = 0 on ΓDir
× (0, T ),
d × (0, T ) and σ(u)ν = 0 on Γd
(4.3b)
0 ≥ ν T u ⊥ ν T (σ(u)ν) ≤ 0 and τ T (σ(u)ν) = 0 on ΓSig
d × (0, T ),
where ν and τ denote the outward unit normal and tangential vectors and (·)T is the transpose of
a tensor. Here the orthogonality has the natural meaning; namely if we have enough regularity, it
means that the product ν T u(ν T σ(u)ν) vanishes almost everywhere at the contact boundary. If
it is not the case, the above inequality is integrated on an appropriate set of test functions, leading
to a weak formulation for the unilateral condition. Let us describe now the functional hypotheses
on the data; if X is a space of scalar functions, the bold–face notation X denotes the space X d .
For the final result, we require the initial displacement u0 belongs to the H1 (Ωd ) and satisfies
the compatibility conditions, i.e. u0 = 0 on ΓDir
and the initial velocity v 0 belongs to L2 (Ωd ).
d
def
It is convenient to introduce the following notations: V = {u ∈ H1 (Ωd ) : u = 0 a.e. on ΓDir
d },
def
def
2
2
2
T
H = L (Ωd ) and the convex set K = {u ∈ L (0, T ; V) : ut ∈ L (0, T ; H), u(·, t) ∈ V, ν u ≤
0 a.e. on ΓSig
d for a.e. t}. Then the weak formulation associated to (4.1)–(4.3) is given by


find u : [0, T ] → K such that


∫
∫
ρutt ·(v−u) dx + a(u, v−u) ≥
f ·(v−u) dx,
(4.4)

Ω
Ω


 for all v ∈ K,
def ∫
where a(u, v) = Ωd Aijkl εij (u)εkl (v) dx. Observe that existence and uniqueness of the solution
to (4.1)–(4.3) is still an open question. However, we have exhibited an approximated solution
associated to (4.4) combining the finite element and mass redistribution methods. Furthermore
note that the energy associated to (4.4) is given by
∫
∫
)
1 (
2
2
E(t) =
|ut (x, t)| +|∇u(x, t)| dx −
f ·ut dx.
2 Ω
Ω
110
4.2. The finite element approximation of problem
Introducing the Lagrange multiplier λ, then the inequality (4.4) is also formally equivalent to:


Find u : [0, T ] → V and λ : [0,

∫ T ] → Λν such that for all v ∈ V



hρutt (t), vi 0 + a(u, v) =
f .v dx + hλν (t), vν i 0
a.e. t ∈ [0, T ],
V ,V
Xν ,Xν
Ω


hµν −λν (t), uν iX0ν ,Xν ≥ 0 and λν (t) ≤ 0,




u(·, 0) = u0 and ut (·, 0) = v 0 ,
for all µν ∈ Λν a.e. t ∈ [0, T ],
where Xν = {ω ∈ L2 (ΓC ) : ∃v ∈ V and ω = vν a.e. on ΓC }, Λν = {µν ∈ X0ν : hµν , vν iX0ν ,Xν ≥
0, for vν ∈ V, vν ≤ 0 a.e. on ΓC } and (uν = u·ν. Here h·, ·iV0 ,V , h·, ·iX0ν ,Xν are the duality
pairing between V0 and V, X0ν and Xν , respectively.
def
def
The present note is organized as follows. A space semi–discretization based on a redistribution of mass is introduced in Section 4.2.1. Then two time–integration methods, namely the
Crank–Nicolson and backward Euler methods are presented in Section 4.2.2. In Section 4.3, the
comparison between the convergence rates for the solution and energy evolution obtained with
and without the mass redistribution highlighted the efficiency of mass redistribution method.
4.2
4.2.1
The finite element approximation of problem
The semi–discretization in space
We introduce the semi–discrete problem in space associated to (4.4) by using a Lagrange finite
element method defined on Ωd . To this aim, we introduce the following notations: Let Th be a
def
regular mesh of Ωd , Vh = {vh ∈ C0 (Ωd ) : vh is piecewise linear over each Kl ∈ Th \ ΓDir
d , vh =
Dir
0 on Γd }, nd and nc denote the number of degree of freedom and the number of nodes on
ΓSig
d , respectively, and ai , i = 1, . . . , n, denotes the finite element nodes. The basis of Vh is
defined using the set of shape functions ϕi ∈ Vh , i = 1, . . . , nd . Then the vector of degree
of freedom of the finite element field uh (x,∫t) denoted by Uh = (u1 (t), . . . , und (t))T∫ such that
∑ d
uh (x, t) = nj=1
uj (t)ϕj (x). Let Mij = ρ Ω ϕi ·ϕj dx, Sij = a(ϕi , ϕj ) and Fi = Ω f ·ϕi dx
denote the components of the mass and stiffness matrices and the external forces, i, j = 1, . . . , nd ,
def
def
respectively. Furthermore let B = (ν1 , . . . , νnc )T and λ = (λ1 , . . . , λnc )T . According to these
notations, a finite element semi–discretization of (4.4) with nodal approximation of the contact
condition is given by


find Uh : [0, T ] → Rnd and λ : [0, T ] → Rnc such that



MU + SU = F + BT λ for a.e. t ∈ [0, T ],
h,tt
h
(4.5)
T

0 ≥ νi Uh ⊥ λi ≤ 0 for all i ∈ Ic and for a.e. t ∈ [0, T ],




Uh (0) = U0h and Uh,t (0) = Vh0 ,
T
T
T
where Ic = {i : ai ∈ ΓSig
d } and such that (ν uh )(ai , t) = (νi Uh )(t) for all i ∈ Ic with νi νj = δi,j
and kνi k = 1, k · k being the Euclidean norm. Note that the Lagrange multipliers are indeed the
nodal contact equivalent forces. The discrete energy associated to problem (4.5) is given by
def
(4.6)
1
T
Eh (t) = (UT
MUh,t +UT
h SUh −Uh,t F)(t).
2 h,t
111
Chapter 4. Numerical study of convergence of the mass redistribution method for
elastodynamic contact problems
The multiplicity of the solution to Problem (4.5) allow us to conclude that the considered problem
is ill–posed, the reader is referred to [94, 85, 84]) for further details. An alternative approach
to the standard discretization presented above is to consider the mass redistributions method
which consists
to replace the mass matrix M by a modified mass matrix Mmod defined by
∫
def
Sig
Mijmod = ρ Ωmod ϕi ·ϕj dx with Ωmod
d,h = {Kl : Kl ∈ Th , Kl ∩ Γd = ∅} for all i, j = 1, . . . , nd .
d,h
Note that ker(Mmod ) = N where N = span{ν1 , . . . , νnc } denotes the space spanned by νi for
⊥
N⊥
i ∈ Ic . Thus employing the identity Uh (t) = UN
is the orthogonal
h (t) + Uh (t) where N
mod
complement of N , and replacing the mass matrix M by the modified one M
in (4.5), we get

⊥

: [0, T ] → N , UN
: [0, T ] → N ⊥ and λ : [0, T ] → Rnc such that
find UN

h
h


Mmod UN ⊥ + S(UN + U N ⊥ ) = F + BT λ for a.e. t ∈ [0, T ],
h,tt
h
h
(4.7)
T
N

0 ≥ νi Uh ⊥ λi ≤ 0 for all i ∈ Ic and for a.e. t ∈ [0, T ],




Uh (0) = U0h and Uh,t (0) = Vh0 .
def
Existence and uniqueness of solution associated to (4.7) as well as the energy balance have
been established in [88, 84]. Furthermore, the convergence of the mass redistribution method
is established in the one–dimensional case in [88] confirming the observations already made in
[111]. Note that the convergence of the mass redistribution is still an open problem in the higher
dimension space.
4.2.2
The time integration methods
We introduce now the time discretization. In order to fix the notations, let the time interval [0, T ]
be divided by n + 1 discrete time–points such that 0 = t0 < t1 < . . . < tn = T . Furthermore the
discrete quantities Unh , Unh,t , Unh,tt and λn are assumed to be given by algorithmic approximations
of the displacement Uh (tn ), the velocity Uh,t (tn ), the acceleration Uh,tt (tn ) and the Lagrange
multiplier λ(tn ), respectively. Some time–stepping schemes allowing to obtain an approximated
solution to Problem (4.5) are introduced below and their efficiency is discussed and analyzed in
the next section.
One of the most popular method in the community of computational mechanics is the Crank–
Nicolson method. This method is second–order consistent and unconditionally stable in the
unconstrained case. Furthermore, the total energy of the discrete evolution is preserved, for the
purely elastic case, see [112]. However, the situation is quite different in the contact constraints
case, the order of accuracy is degraded, for further details, the reader is referred to [120, 121, 122].
Notice that the discrete evolution associated to (4.5) can be described by the finite difference
equations:


find Un+1
: [0, T ] → Rnd and λn+1 : [0, T ] → Rnc such that

h


2


n+1

n + ∆tUn + ∆t (Un +Un+1 ),

U
=
U

h,tt
h,t
h
h
h,tt

4
∆t
n+1
n
(4.8)
n
Un+1

h,t = Uh,t + 2 (Uh,tt +Uh,tt ),



n+1


MUn+1
= Fn+1 + BT λn+1 ,

h,tt + SUh


0 ≥ ν T Un+1 ⊥ λn+1 ≤ 0 for all i ∈ I ,
i
h
c
i
112
4.3. Numerical experiments
where ∆t is a given times step and U0h , U0h,t and λ0 are given. Therefore we observe that (4.8)
leads to the following algorithm:
(4.9)


find Un+1
: [0, T ] → Rnd and λn+1 : [0, T ] → Rnc such that

h
(
)
4M n
4M
n
n+1
+S Un+1
=
(Uh +∆tUn+1
+ BT λn+1 ,
h
h,t ) + MUh,tt + F
2
2

∆t
∆t


⊥ λn+1
≤ 0 for all i ∈ Ic ,
0 ≥ νiT Un+1
i
h
The energy evolution is defined by ∆Ehn = Ehn+1 − Ehn , where Ehn is supposed to be given by an
algorithmic approximation of the energy Eh (tn ). Thus the energy evolution associated to (4.8)
n T T
n
n+1 ).
by using the Crank–Nicolson method is ∆Ehn = − 21 (Un+1
h −Uh ) B (λ +λ
def
Another time integration method to approach the semi–discrete problem (4.5) consists to use
the backward Euler method which gives us the following discrete evolution associated to (4.5):


find Un+1
: [0, T ] → Rnd and λn+1 : [0, T ] → Rnc such that

h



n+1
n+1
n


Uh = Uh + ∆tUh,t ,
(4.10)
n+1
n
Un+1
h,t = Uh,t + ∆tUh,tt ,


n+1


MUn+1
= Fn+1 + BT λn+1 ,

h,tt + SUh



0 ≥ νiT Un+1
⊥ λn+1
≤ 0 for all i ∈ Ic ,
i
h
where U0h , U0h,t and λ0 are given. Therefore notice that (4.10) enables us to infer the following
algorithm:
(4.11)


find Un+1
: [0, T ] → Rnd and λn+1 : [0, T ] → Rnc such that

h
(
)
4M
4M n
n+1
+S Un+1
(Uh +∆tUn+1
=
+ BT λn+1 ,
h
h,t ) + F
2
2

∆t
∆t


0 ≥ νiT Un+1
⊥ λn+1
≤ 0 for all i ∈ Ic ,
i
h
n T
In this case, the energy evolution associated to (4.10) reads as ∆Ehn = − 12 (Un+1
h,t − Uh,t ) M
n+1
n+1
n+1
n+1
1
n
n
T
n
n
T
T
n+1
(Uh,t −Uh,t )− 2 (Uh −Uh ) S(Uh −Uh )−∆t(Uh −Uh ) B λ
. The reader is referred
to [111, Appendix] where the energy evolutions associated to the Crank–Nicolson and backward
Euler methods are justified. Note that Uh,tt does not appear on the right hand side of (4.11)
which is not the case in (4.9) implying that the backward Euler method is dissipative and stable.
4.3
Numerical experiments
The parameters used in the numerical simulations are Ω2 = [0, 0.1] × [0, 1], Ω3 = [0, 0.1] ×
[0, 0.1] × [0, 1], ρ = 1, the Lamé parameters λ = 0.25 and µ = 0.5, u(x1 , 1, t) = 0.2, x1 ∈ [0, 0.1]
for dimension two and u(x1 , x2 , 1, t) = 0.2, (x1 , x2 ) ∈ [0, 0.1] × [0, 0.1] for dimension three. The
bar is undeformed at t = 0 ant it is located at the distance of 0.2 from a rigid obstacle. The bar
starts to move toward the rigid obstacle with an initial velocity equal to −0.5 (vertical direction)
and without any external forces. Note that we used the square and cubic meshes with square
and cube elements for two and three dimensional case, respectively.
113
Chapter 4. Numerical study of convergence of the mass redistribution method for
elastodynamic contact problems
The Crank–Nicolson and backward Euler methods are used below to compute (Unh , λn ).
These computations are performed for the standard mass matrix M as well as for the modified
mass matrix M = Mmod . Contrary to the one–dimensional elastodynamic contact problem
treated in [111], we were not able here to exhibit a true solution associated to (4.1)–(4.3).
However the reference solution denoted by (U, λ) can be approximated by (Unh , λn ) by taking
∆t and ∆x very small. Therefore (U, λ) plays the role of an explicit solution in the numerical
simulations presented below. Similarly the energy E is assumed to be equal to Ehn for ∆t and ∆x
chosen sufficiently small.
1
For the 2D convergence curves, the following parameters are used: ∆x1 = ∆x2 = 640
and
−4
∆t = 78125 × 10 , to get the reference solution (U, λ) and the energy E. Therefore the
approximated solution (Unh , λn ) and energy Ehn are obtained starting by the parameters ∆x1 =
1
∆x2 = 20
and ∆t = 25 × 10−2 which are successively divided by 2.
Due to the rapidly growing number of degree of freedom in dimension three, we were only able
1
to compute a reference solution (U, λ) and the energy E with the parameters ∆x1 = ∆x2 = 160
,
1
−3
∆x3 = 32 and ∆t = 625 × 10 . The convergence curves have only three points where the first
point is obtained with the parameters ∆x1 = ∆x2 = 0.1, ∆x3 = 0.5 and ∆t = 0.1 and the others
are obtained by the parameters divided by 2. We may remark that for the Lagrange multipliers,
the convergence does not take place because the parameters are not small enough. The numerical
simulations were performed by employing the finite element library Getfem++ (see [87]).
4.3.1
The Crank–Nicolson scheme
We focus first on the two–dimensional elastodynamic contact problem for which an approximated
solution at the contact nodes as well as some convergence curves were obtained. More precisely,
some spurious oscillations can be observed for the normal displacement at the contact nodes
and for the Lagrange multiplier, when using the standard mass matrix after the contact takes
place, see Figure 4.1. These oscillations do not exist in the case where the standard mass matrix
is replaced by the modified one, see Figure 4.1. Note that the scaling in space is different for
the Lagrange multiplier evaluated by using standard and modified mass matrices. The scheme
with the standard mass matrix is unstable with a rapidly growing energy while the one with the
modified mass matrix is almost conservative as the space ∆x and time ∆t steps become small.
The error curves for (Unh , λn ) in Figure 4.3 highlight the convergence of both the displacement
and the Lagrange multipliers in L2 (0, T ; V) and L2 (0, T ; ΓC ) respectively, which is not the case
for the standard mass matrix. while the displacement in the norm Lp (0, T ; H), p = 2, ∞,
converges in the both cases. The same condition can be drawn for the energy Ehn in Lp (0, T )
norm for p = 2, +∞. The convergence curves in 3D case are represented on the Figure 4.4.
the numerically convergence are very similar to the two–dimensional case except the fact that
there is no clear convergence of the displacement in the L2 (0, T ; V) norm with the modified mass
matrix. The evolution of the von Mises stress in a three–dimensional bar by using the standard
and modified mass matrices is represented in Figures 4.5 and 4.6. As in the two–dimensional
case, some spurious oscillations can be observed in a neighborhood of the contact nodes in the
case where the standard matrix is considered. Then these oscillations propagated along the bar
(see Figure 4.5). In the case where the standard mass matrix is replaced by the modified one,
114
4.3. Numerical experiments
these oscillations do not exist anymore (see Figure 4.6).
0.08
0.06
0.04
0.02
Space
0 0
2
4
6
8
0.08
10
0.06
0.04
0.02
Space
Time
(a)
2
0 0
4
6
8
10
Time
(b)
(c)
(d)
Figure 4.1: Approximated solutions obtained by using the standard (left) and modified (right)
mass matrices in the contact node with Crank–Nicolson scheme.
Figure 4.2: Energy associated to the approximated solution for the standard (left) and modified
(right) mass matrices with Crank–Nicolson scheme.
115
Chapter 4. Numerical study of convergence of the mass redistribution method for
elastodynamic contact problems
(a) kUn
h −UkL∞ (0,T ;H)
(b) kUn
h −UkL2 (0,T ;H)
(c) kUn
h −UkL2 (0,T ;V)
(d) kλn −λkL2 (0,T ;ΓC )
(e) kEh −EkL∞ (0,T )
(f) kEh −EkL2 (0,T )
Figure 4.3: Comparison of the convergence curves obtained by the modified and standard mass
matrices with square elements and ∆x
∆t = 2 for 2D case by using Crank–Nicolson scheme.
116
4.3. Numerical experiments
(a) kUn
h −UkL∞ (0,T ;H)
(b) kUn
h −UkL2 (0,T ;H)
(c) kUn
h −UkL2 (0,T ;V)
(d) kEh −EkL∞ (0,T )
(e) kEh −EkL2 (0,T )
Figure 4.4: Comparison of the convergence curves obtained by the modified and standard mass
matrices with square elements and ∆x
∆t = 2 for 3D case by using Crank–Nicolson scheme
.
117
Chapter 4. Numerical study of convergence of the mass redistribution method for
elastodynamic contact problems
(a) t=0
(b) t=0.2
(c) t=0.4
(d) t=0.6
(e) t=0.8
(f) t=1
(g) t=1.5
(h) t=2
(i) t=2.5
(j) t=3
(k) t=3.5
(l) t=3.75
(m) t=4
(n) t=4.5
(o) t=5
(p) t=5.5
Figure 4.5: Von Mises stress evolution of the deformed bar after the contact by using Crank–
Nicolson scheme and with the standard mass matrix (∆x = 0.02 and ∆t = 0.01).
118
4.3. Numerical experiments
(a) t=0
(b) t=0.2
(c) t=0.4
(d) t=0.6
(e) t=0.8
(f) t=1
(g) t=1.5
(h) t=2
(i) t=2.5
(j) t=3
(k) t=3.5
(l) t=3.75
(m) t=4
(n) t=4.5
(o) t=5
(p) t=5.5
Figure 4.6: Von Mises stress evolution of the deformed bar after the contact by using Crank–
Nicolson scheme and with the modified mass matrix (∆x = 0.02 and ∆t = 0.01).
119
Chapter 4. Numerical study of convergence of the mass redistribution method for
elastodynamic contact problems
4.3.2
The backward Euler method
The normal displacement at the contact nodes and the Lagrange multiplier obtained by using
the standard mass as well as modified mass matrices do not oscillate (see Figure 4.7) as it can
be observed for the Crank–Nicolson method. This is due to the unconditional stability and
dissipativity of Euler method. Furthermore, the error curves converge for both methods (see
Figures 4.9 and 4.10) and the energy are almost the same (see Figure 4.8).
0.08
0.06
0.04
0.02
Space
0
0
2
4
6
8
0.08
10
0.06
0.04
0.02
Space
Time
0
0
2
4
6
8
10
Time
Figure 4.7: Approximated solutions obtained by using the standard (left) and modified (right)
mass matrices in the contact node with the backward Euler method.
Figure 4.8: Energy associated to the approximated solution for the standard (left) and modified
(right) mass matrices with the backward Euler method.
120
4.3. Numerical experiments
(a) kUn
h −UkL∞ (0,T ;H)
(b) kUn
h −UkL2 (0,T ;H)
(c) kUn
h −UkL2 (0,T ;V)
(d) kλn −λkL2 (0,T )
(e) kEh −EkL∞ (0,T )
(f) kEh −EkL2 (0,T )
Figure 4.9: Comparison of the convergence curves obtained by the modified and standard mass
matrices with square elements ∆x
∆t = 2 for 2D case by using the backward Euler method
.
121
Chapter 4. Numerical study of convergence of the mass redistribution method for
elastodynamic contact problems
(a) kUn
h −UkL∞ (0,T ;H)
(b) kUn
h −UkL2 (0,T ;H)
(c) kUn
h −UkL2 (0,T ;V)
(d) kEh −EkL∞ (0,T )
(e) kEh −EkL2 (0,T )
Figure 4.10: Comparison of the convergence curves obtained by the modified and standard mass
matrices with square elements and ∆x
∆t = 2 for 3D case by using the backward Euler method
.
122
4.4. Some complements
4.4
Some complements
The straightforward application of Newmark method to contact Problem (4.5) can be written as
follows:

n+1
nd
n+1 : [0, T ] → Rnc such that


Find Uh : [0, T ] → R ( and λ

)

1

n+1
2 n
2 n+1
n
n


Uh = Uh + ∆tUh,t + 2 −β ∆t Uh,tt + β∆t Uh,tt ,
n+1
n
n
(4.12)
Un+1
h,t = Uh,t + (1−γ)∆tUh,tt + γ∆tUh,tt ,



n+1

MUn+1
= Fn+1 + BT λn+1


h,tt + SUh


0 ≥ ν T Un+1 ⊥ λn+1 ≤ 0 for all i ∈ I ,
c
i
i
h
where ∆t is a given times step and U0h , U0h,t and λ0 are given. The Problem (4.12) leads to the
following algorithm:


Find Un+1 : [0, T ] → Rnd and λn+1 : [0, T ] → Rnc such that


( M h )
1−2β
M
MUnh,tt + Fn+1 + BT λn+1 ,
+S Un+1
(Unh +∆tUnh,t ) +
=
h
2
2

β∆t
β∆t
2β


0 ≥ ν T Un+1 ⊥ λn+1 ≤ 0 for all i ∈ I .
c
i
i
h
This method is unconditionally stable for linear elastodynamic problem for γ ≥ 21 and β ≥
1 1
2
4 ( 2 +γ) (see [120, 121]). The energy evolution associated to (4.12) for F independent of time
is given by
(
γ) n T
∆Ehn = (1−2γ)(Vhn )T SVhn + ∆t β− (Vh,t
) SVhn
2
(
(4.13)
γ) n T T n
+ ∆t β− (Vh,t
) (B Λ ) − ∆t(Vhn )T (BT Λnγ ),
2
where Vhn = Un+1
− Unh , Λn = λn+1 − λn and Λnγ = (1−γ)λn + γλn+1 (see [111]).
h
def
4.4.1
def
def
The Newmark method I
We consider now the Newmark method with the couple (β, γ) = ( 12 , 1). This method is dissipative
and consequently is stable, since from (4.13), we have
∆Ehn = −(Vhn )T SVhn − ∆t(Vhn )T (BT Λn1 ) < 0.
We observe that the normal displacement at the contact node does not oscillate by using the
both standard and modified mass matrices while there are some oscillations for the Lagrange
multiplier, when the standard matrix is used. These oscillations are reduced in the case of the
modified matrix (see Figure 4.11). Note that the energies for both matrices are decreasing (see
Figure 4.12). The convergence curves in 2D are represented on the Figure 4.13. We observe that
the Lagrange multiplier in the L2 (0, T ; H) norm does not converge when the standard matrix is
used. In the other cases, we have the convergence with almost the same rates. In 3D, there is
no clear convergence in L2 (0, T ; V) norm (see Figure 4.14).
123
Chapter 4. Numerical study of convergence of the mass redistribution method for
elastodynamic contact problems
0.08
0.06
0.04
0.02
Space
0
0
2
4
6
8
0.08
10
0.06
0.04
0.02
Space
Time
0
0
2
4
6
8
10
Time
Figure 4.11: Approximated solutions obtained by using the standard (left) and modified (right)
mass matrices in the contact node with the Newmark method I.
Figure 4.12: Energy associated to the approximated solution for the standard (left) and modified
(right) mass matrices with the Newmark method I.
124
4.4. Some complements
(a) kUhn − U kL∞ (0,T ;H)
(b) kUhn − U kL2 (0,T ;H)
(c) kUhn − U kL2 (0,T ;V)
(d) kλn − λkL2 (0,T ;H)
(e) kEh − EkL∞ (0,T )
(f) kEh − EkL2 (0,T )
Figure 4.13: Comparison of the convergence curves obtained by the modified and standard mass
matrices with square elements ∆x
∆t = 2 for 2D case by using the Newmark method I.
125
Chapter 4. Numerical study of convergence of the mass redistribution method for
elastodynamic contact problems
(a) kUn
h −UkL∞ (0,T ;H)
(b) kUn
h −UkL2 (0,T ;H)
(c) kUn
h −UkL2 (0,T ;V)
(d) kEh −EkL∞ (0,T )
(e) kEh −EkL2 (0,T )
Figure 4.14: Comparison of the convergence curves obtained by the modified and standard mass
matrices with square elements and ∆x
∆t = 2 for 3D case by using the Newmark method I
.
126
4.4. Some complements
4.4.2
The Newmark method II
We deal now with the Newmark method with the couple (β, γ) = ( 12 , 12 ). We observe some
oscillations of the Lagrange multiplier in Figure 4.15 and also a small instability of the energy
evolution in Figure 4.16. When the standard matrix is replaced by the modified matrix, these
oscillations of Lagrange multiplier are reduced. Moreover the energy is almost conserved (see
Figure 4.16). The convergence curves in dimension two illustrate that the Lagrange multiplier
does not converge in L2 (0, T ; H) norm when the standard mass matrix is used (see Figure 4.17). In
dimension three, the displacement in the L2 (0, T ; V) norm does not converge when the standard
matrix is used (see Figure 4.18).
0.08
0.06
0.04
0.02
Space
0
0
2
4
6
8
0.08
10
0.06
0.04
0.02
Space
Time
0
0
2
4
6
8
10
Time
Figure 4.15: Approximated solutions obtained by using the standard (left) and modified (right)
mass matrices in the contact node with the Newmark method II.
Figure 4.16: Energy associated to the approximated solution for the standard (left) and modified
(right) mass matrices with the Newmark method II.
127
Chapter 4. Numerical study of convergence of the mass redistribution method for
elastodynamic contact problems
(a) kUhn − U kL∞ (0,T ;H)
(b) kUhn − U kL2 (0,T ;H)
(c) kUhn − U kL2 (0,T ;V)
(d) kλn − λkL2 (0,T ;H)
(e) kEh − EkL∞ (0,T )
(f) kEh − EkL2 (0,T )
Figure 4.17: Comparison of the convergence curves obtained by the modified and standard mass
matrices with square elements ∆x
∆t = 2 for 2D case by using the Newmark method II.
128
4.4. Some complements
(a) kUn
h −UkL∞ (0,T ;H)
(b) kUn
h −UkL2 (0,T ;H)
(c) kUn
h −UkL2 (0,T ;V)
(d) kEh −EkL∞ (0,T )
(e) kEh −EkL2 (0,T )
Figure 4.18: Comparison of the convergence curves obtained by the modified and standard mass
matrices with square elements and ∆x
∆t = 2 for 3D case by using the Newmark method II
.
129
Chapter 4. Numerical study of convergence of the mass redistribution method for
elastodynamic contact problems
4.4.3
The convergence rates
We summarize here the convergence rates of different time integration methods in the two and
three–dimensional cases. In two dimensional case, we observe that by using the second order
methods, namely the Crank–Nicolson and the Newmark method II, the convergence rates of
displacement in Lp (0, T ; H) for p = 2, ∞ are still 12 for the standard mass matrix. These convergence rates are improved for the modified mass matrix. Moreover the convergence of the
Lagrange multipliers by using the second order methods for standard mass matrix do not occur.
While for the modified mass matrix, they converge with the rate approximately equal to 12 . The
use of the stable one order methods, namely the backward Euler and the Newmark method I,
gives lower rates of convergence. Furthermore for these methods, there are no substantial differences between the convergence rates obtained by using the standard as well as modified mass
matrices.
Scheme
Crank–Nicolson
Newmark I (β, γ) = ( 12 , 1)
Newmark II (β, γ) = ( 12 , 12 )
Backward Euler
Method
STD
MOD
STD
MOD
STD
MOD
STD
MOD
kUhn −ukL∞ (0,T ;H)
0.5732
0.9009
0.6007
0.4711
0.5757
0.8754
0.6129
0.5650
kUhn −ukL2 (0,T ;H)
0.5756
0.9536
0.5672
0.5058
0.7893
0.8496
0.6007
0.5259
kUhn −ukL2 (0,T ;V)
-0.4453
0.3419
0.2563
0.2529
0.3133
0.3160
0.1953
0.1919
Table 4.1: Convergence rates of displacement for the two dimensional case
Scheme
Crank–Nicolson
Newmark I (β, γ) = ( 12 , 1)
Newmark II (β, γ) = ( 12 , 12 )
Backward Euler
Method
STD
MOD
STD
MOD
STD
MOD
STD
MOD
kλh −λkL2 (0,T ;H)
-0.2310
0.4416
0.0081
0.4782
-0.0186
0.5766
0.4377
0.4379
kEh −EkL∞ (0,T )
-0.4855
1.1086
0.4710
0.4621
1.1853
1.0864
0.4744
0.4631
kEh −EkL2 (0,T )
-0.4623
1.1386
0.4621
0.4888
1.1660
1.1700
0.4896
0.5005
Table 4.2: Convergence rates of Lagrange multiplier and energy for the two dimensional case
Concerning the three–dimensional case, since the reference solution is computed numerically,
therefore the convergence curves are obtained by limited space and time steps. However theses
convergence curves confirm the efficiency of the mass modified matrix. We may conclude that
the second order time integration methods are less stable than the one order methods presented.
Indeed, some spurious oscillations and unfavorable behavior for the energy of system are highlighted. The mass redistribution method reduces these oscillations for the displacement and
the contact force. Furthermore, better behaviors for the energy are obtained and the convergence rates are improved for the the displacement and the contact force. We may conclude that
130
4.4. Some complements
the second order time integration methods together with the mass redistribution method is a
numerically robust method for approximating the elastodynamic contact problems.
Scheme
Method
Crank–Nicolson
Newmark I (β, γ) = ( 12 , 1)
Newmark II (β, γ) = ( 12 , 21 )
Backward Euler
STD
MOD
STD
MOD
STD
MOD
STD
MOD
kUhn −ukL∞ (0,T ;H)
0.2958
1.2600
0.2563
0.8113
0.5205
1.2521
0.1535
0.7356
kUhn −ukL2 (0,T ;H)
0.3731
1.2482
0.3417
0.8563
0.4212
1.2203
0.1701
0.7507
kUhn −ukL2 (0,T ;V)
-0.2951
0.4087
-0.2625
0.4073
-0.1859
0.3851
-0.3097
0.3884
Table 4.3: Convergence rates of displacement for the three dimensional case
Scheme
Crank–Nicolson
Newmark I (β, γ) = ( 12 , 1)
Newmark II (β, γ) = ( 12 , 12 )
Backward Euler
Method
STD
MOD
STD
MOD
STD
MOD
STD
MOD
kEh −EkL∞ (0,T )
-0.8684
1.1567
0.5369
0.7720
1.0310
1.0910
0.5079
0.5687
kEh −EkL2 (0,T )
-1.2925
1.1843
0.7181
0.8941
1.0670
1.1460
0.6647
0.8684
Table 4.4: Convergence rates of energy for the three dimensional case
Appendix F
As a complement, let us consider here the case when the bar is dropped toward the rigid obstacle (here ΓD = ∅). The numerical three–dimensional simulations are performed by using the
standard and the modified mass matrices and with the Crank–Nicolson scheme. The parameters
used here are Ω3 = [0, 0.1] × [0, 0.1] × [0, 1], ρ = 1, the Lamé parameters λ = 0.25 and µ = 0.5.
We supposed that the Neumann conditions vanishes. The bar is undeformed at t = 0 ant it
is located at the distance of 0.2 from a rigid obstacle. The bar starts to fall down to the rigid
obstacle with an initial velocity equal to −0.5 (vertical direction). Note that any external forces
are applied. Here the cubic elements with ∆x = 0.02 are used in numerical experiments. With
the standard mass matrix we may observe clearly some oscillations on the body of the bar after
the impact takes place. Whereas with the modified mass matrix, the oscillations vanish (see
Figures 4.19 and 4.19). Note that the convergence rates for this case validate the convergence of
the mass redistribution method.
131
Chapter 4. Numerical study of convergence of the mass redistribution method for
elastodynamic contact problems
(a) t=0
(b) t=0.2
(c) t=0.4
(d) t=0.6
(e) t=0.8
(f) t=1
(g) t=1.5
(h) t=2
(i) t=2.5
(j) t=3
(k) t=3.5
(l) t=3.75
(m) t=4
(n) t=4.5
(o) t=5
(p) t=5.5
Figure 4.19: Von Mises stress evolution of the deformed bar after the contact by using Crank–
Nicolson scheme and with the standard mass matrix (∆x = 0.02 and ∆t = 0.01).
132
4.4. Some complements
(a) t=0
(b) t=0.2
(c) t=0.4
(d) t=0.6
(e) t=0.8
(f) t=1
(g) t=1.5
(h) t=2
(i) t=2.5
(j) t=3
(k) t=3.5
(l) t=3.75
(m) t=4
(n) t=4.5
(o) t=5
(p) t=5.5
Figure 4.20: Von Mises stress evolution of the deformed bar after the contact by using Crank–
Nicolson scheme and with the modified mass matrix (∆x = 0.02 and ∆t = 0.01).
133
Bibliography
[1] M. Brinkmeier, U. Nackenhorst, S. Petersen, and O. Von Estorff. A finite element approach
for the simulation of tire rolling noise. Journal of Sound and Vibration, 309(1–2):20–39,
2008. (Cited on page 1.)
[2] L. Euler. Sur la diminution de la résistance du frottement. Mem. Acad. Sci. Berlin,
4:133–148, 1748. (Cited on page 2.)
[3] C. A. Coulomb. Théorie des machines simples, en ayant égard au frottement de leurs
parties et la raideur des cordages. Prix double Acad. Sci., 1781. Mémoirs Savants Etrang.
X, 163–332, 1785 (reprinted by Bachelier, Paris 1809). (Cited on page 3.)
[4] H. Hertz. Uber die beruhrung fester elastischer korper. Journal fur die Reine und Angewandte Mathematik, 29:156–171, 1882. (Cited on page 3.)
[5] L. A. Galin. Contact problems in the theory of elasticity. Translated from the Russia by H.
Moss, North Carolina State College, Depatement of Mathematics, 1961. (Cited on page 4.)
[6] S. P. Timoshenko and J. N. Goodier. Theory of Elasticity. McGraw-Hill, New York, 1970.
(Cited on page 4.)
[7] G. M. L. Gladwell. Contact problems in the classical theory of elasticity. Alphen aden Rijn:
Sijthoff and Noordhoff, 1980. (Cited on page 4.)
[8] K. L. Johnson. Contact Mechanics. Cambridge University Press, 1985. (Cited on page 4.)
[9] D. Doyen, A. Ern, and S. Piperno. Time–integration schemes for the finite element dynamic
Signorini problem. SIAM J. Sci. Comput., 33(1):223–249, 2011. (Cited on pages 4 and 42.)
[10] A. Signorini. Sopra alcune questioni di statica dei sistemi continui. Ann. Scuola Norm.
Sup. Pisa Cl. Sci. (2), 2(2):231–251, 1933. (Cited on page 4.)
[11] G. Fichera. Problemi elastostatici con vincoli unilaterali: il problema di signorini con
ambique condizioni al contorno. Mem. Acc. Naz. Lincei, VIII (7):91–140, 1964. (Cited on
page 4.)
[12] J. A. C. Martins and J. T. Oden. Existence and uniqueness results for dynamic contact
problems with nonlinear normal and friction interface laws. Nonlinear Anal., 11(3):407–
428, 1987. (Cited on page 4.)
[13] A. Klarbring, A. Mikelić, and M. Shillor. Frictional contact problems with normal compliance. Internat. J. Engrg. Sci., 26(8):811–832, 1988. (Cited on page 4.)
[14] A. Klarbring, A. Mikelić, and M. Shillor. A global existence result for the quasistatic
frictional contact problem with normal compliance. In Unilateral problems in structural
analysis, IV (Capri, 1989), volume 101 of Internat. Ser. Numer. Math., pages 85–111.
Birkhäuser, Basel, 1991. (Cited on page 4.)
Bibliography
[15] L. Andersson. A quasistatic frictional problem with a normal compliance penalization
term. Nonlinear Anal., 37(6, Ser. A: Theory Methods):689–705, 1999. (Cited on page 4.)
[16] L. Andersson. A quasistatic frictional problem with normal compliance. Nonlinear Anal.,
16(4):347–369, 1991. (Cited on page 4.)
[17] C. Eck and J. Jaru˘sek. Existence results for the static contact problem with Coulomb
friction. Math. Models Methods Appl. Sci, 8 (3):445–468, 1998. (Cited on page 4.)
[18] C. Eck and J. Jaru˘sek. Existence results for the semicoercive static contact problem with
coulomb friction. Nonlinear Anal., Theory Methods Appl, 42 A(6):961–976, 2000. (Cited
on page 4.)
[19] I. Bock and J. Jaru˘sek. On hyperbolic contact problems. Tatra Mt. Math. Publ., 43:25–40,
2009. (Cited on page 4.)
[20] M. Sofonea, W. Han, and M. Barboteu. Analysis of a viscoelastic contact problem with
multivalued normal compliance and unilateral constraint. Comput. Methods Appl. Mech.
Engrg., 264:12–22, 2013. (Cited on page 4.)
[21] J. A. C. Martins, F. M. F., and F. Gastaldi. On a quasi static contact problem in viscoelsaticity with normal compliance. Instituto Di Analisi Numerica, 1993. (Cited on page 4.)
[22] R. Dautray and J. L. Lions. Analyse mathématique et calcul numérique pour les sciences
et les techniques. Vol. 8. INSTN: Collection Enseignement. [INSTN: Teaching Collection].
Masson, Paris, 1988. Évolution: numérique, transport. [Evolution: numerical methods,
transport], Reprint of the 1985 edition. (Cited on pages 4, 5, 52 and 53.)
[23] J. Ne˘cas, J. Jaru˘sek, and J. Haslinger. On the solution of the variational inequality to the
signorini problem with small friction. Boll. Un. Mat. Ital., 5 (17–B):796–811, 1980. (Cited
on page 4.)
[24] P. D. Panagiotopoulos. A nonlinear programming approach to the unilateral contact-, and
friction–boundary value problem in the theory of elasticity. Ingenieur–Archiv, 44 ,Issue
6:421–432, 1975. (Cited on page 4.)
[25] S. Agmon, A. Douglis, and L. Nirenberg. Estimates near boundary for solutions of elliptic
partial differential equations satisfying general boundary conditions. I. Comm. Pure Appl.
Math, 12:623–727, 1959. (Cited on page 4.)
[26] G. Fichera. Existence theorems in elasticity. boundary value problems of elasticity with
unilateral constraints. Springer–Verlag, Berlin, 2:374–424, 1972. (Cited on page 4.)
[27] J. Jaru˘sek. Contact problems with bounded friction. Coercive case. Czechoslovak Math J,
33 (108):237–261, 1983. (Cited on page 4.)
[28] J. Jaru˘sek. Contact problems with bounded friction. Semicoercive case. Czechoslovak Math.
J, 34 (109):619–629, 1984. (Cited on page 4.)
136
Bibliography
[29] M. Cocu. Existence of solutions of Signorini problems with friction. Internat. J. Engrg.
Sci., 22(5):567–575, 1984. (Cited on page 4.)
[30] P. Hild. An example of nonuniqueness for the continuous static unilateral contact model
with coulomb friction. C. R. Acad. Sci. Paris, 337:685–688, 2003. (Cited on page 4.)
[31] P. Hild. Non–unique slipping in the Coulomb friction model in two-dimensional linear
elasticity. Quart J. Mech. Appl. Math, 57:235–245, 2004. (Cited on page 4.)
[32] P. Ballard. Steady sliding frictional contact problems in linear elasticity. J. Elasticity,
110(1):33–61, 2013. (Cited on page 4.)
[33] P. Ballard. A counter–example to uniqueness in quasi–static elastic contact problems with
small friction. Internat. J. Engrg. Sci., 37(2):163–178, 1999. (Cited on page 4.)
[34] J. A. C. Martins, Manuel D. P. Monteiro Marques, and F. Gastaldi. On an example of
nonexistence of solution to a quasistatic frictional contact problem. European J. Mech. A
Solids, 13(1):113–133, 1994. (Cited on page 4.)
[35] Y. Renard. A uniqueness criterion for the Signorini problem with Coulomb friction. SIAM
J. MATH. ANAL, 38(2):452–467, 2006. (Cited on page 5.)
[36] C. Eck, J. Jaru˘sek, and M. Krbec. Unilateral contact problems : variational methods and
existence theorems. Monographs and textbooks in pure and applied mathematics. Chapman
& Hall, Boca Raton, London, New York, 2005. (Cited on pages 5 and 56.)
[37] L. Amerio and G. Prouse. Study of the motion of a string vibrating against an obstacle.
Rend. Mat, Ser. 6, 8:563–585, 1975. (Cited on page 5.)
[38] L. Amerio. On the motion of a string vibrating through a moving ring with a continuously
variable diameter. Att. Acad. Naz. Lincei Rend, 62:134–142, 1977. (Cited on page 5.)
[39] A. Bamberger and M. Schatzman. New results on the vibrating string with a continuous
obstacle. SIAM Journal on Mathematical Analysis, 14(3):560–595, 1983. (Cited on page 5.)
[40] M. Schatzman. A hyperbolic problem of second order with unilateral constraints: The vibrating string with a concave obstacle. Journal of Mathematical Analysis and Applications,
73(1):138–191, 1980. (Cited on pages 5, 25, 56, 58 and 94.)
[41] M. Schatzman. Un problème hyperbolique du 2ème ordre avec contrainte unilatérale:
la corde vibrante avec obstacle ponctuel. J. Differential Equations, 36(2):295–334, 1980.
(Cited on page 5.)
[42] C. Citrini and C. Marchionna. On the motion of a string vibrating against a gluing obstacle.
Rend. Accad. Naz. Sci. Detta XL, V. Ser., Mem. Mat., 13(1):141–164, 1989. (Cited on
page 5.)
[43] C. Citrini and C. Marchionna. Some unilateral problems for the vibrating string equation.
Eur. J. Mech., A, 8(1):73–85, 1989. (Cited on page 5.)
137
Bibliography
[44] G. Lebeau and M. Schatzman. A wave problem in a half–space with a unilateral constraint
at the boundary. J. Differential Equations, 53(3):309–361, 1984. (Cited on pages 5, 25, 26,
56, 58, 94 and 109.)
[45] J. U. Kim. A boundary thin obstacle problem for a wave equation. Comm. Partial Differential Equations, 14(8-9):1011–1026, 1989. (Cited on pages 5, 25, 56, 58, 94 and 109.)
[46] J. E. Muñoz Rivera and R. Racke. Multidimensional contact problems in thermoelasticity.
SIAM J. Appl. Math, 58(4):1307–1337, 1998. (Cited on page 5.)
[47] J. Jaru˘sek. Contact problems with given time–dependent friction force in linear viscoelasticity. Comment. Math. Univ. Carolin., 31(2):257–262, 1990. (Cited on page 5.)
˘
[48] J. Jaru˘sek, J. Málek, J. Ne˘cas, and V. Sverák.
Variational inequality for a viscous drum
vibrating in the presence of an obstacle. Rend. Mat. Appl. (7), 12(4):943–958 (1993), 1992.
(Cited on page 5.)
[49] J. Jaru˘sek. Solvability of unilateral hyperbolic problems involving viscoelasticity via penalization. SAACM, 3 (2):129–140, 1993. (Cited on page 5.)
[50] J. Jaru˘sek. Dynamic contact problems with given friction for viscoelastic bodies. Czechoslovak Math. J., 46(121)(3):475–487, 1996. (Cited on page 5.)
[51] J. Jaru˘sek and C. Eck. Dynamic contact problems with small Coulomb friction for viscoelastic bodies. Existence of solutions. Math. Models Methods Appl. Sci., 9(1):11–34, 1999.
(Cited on page 5.)
[52] C. Eck and J. Jaru˘sek. Existence of solutions for the dynamic frictional contact problem
of isotropic viscoelastic bodies. Nonlinear Anal., 53(2):157–181, 2003. (Cited on page 5.)
[53] A. Petrov and M. Schatzman. Mathematical results on existence for viscoelastodynamic
problems with unilateral constraints. SIAM J. Math. Anal., 40(5):1882–1904, 2008/09.
(Cited on pages 5 and 56.)
[54] A. Petrov and M. Schatzman. Viscoélastodynamique monodimensionnelle avec conditions
de Signorini. C. R. Acad. Sci. Paris Sér. I Math., 11(334):983–988, 2002. (Cited on page 5.)
[55] A. Petrov and M. Schatzman. A pseudifferential linear complementarity problem related
to one dimensional viscoelastic model with Signorini conditions. Arch. Rat. Mech. Anal.,
to appear, 2014. (Cited on page 5.)
[56] E. A. Wilson and B. Parsons. Finite element analysis of elastic contact problems using
differential displacements. Int. J. Numer. Meth. Engng., 2:387–395, 1970. (Cited on page 5.)
[57] S. K. Chan and I. S. Tuba. A finite element method for contact problems of solid bodies–
Part i. theory and validation. International Journal of Mechanical Sciences, 13(7):615–625,
1971. (Cited on pages 5 and 6.)
[58] T. F. Conry and A. Seireg. A mathematical programming method for design of elastic
bodies in contact. J. Appl. Mech., 38(2):387–392, 1971. (Cited on page 5.)
138
Bibliography
[59] T. J. R. Hughes, R. L. Taylor, J. L. Sackman, A. Curnier, and W. Kanoknukulchai. A
finite element method for a class of contact–impact problems. Computer Methods in Applied
Mechanics and Engineering, 8(3):249–276, 1976. (Cited on pages 6 and 39.)
[60] A. Francavilla and O. C. Zienkiewicz. A note on numerical computation of elastic contact
problems. Int. J. Numer. Meth. Engng, 9:913–924, 1975. (Cited on page 6.)
[61] J. Haslinger and I. Hlaváček. Contact between elastic bodies. II. Finite element analysis.
Apl. Mat., 26(4):263–290, 1981. With a loose Russian summary. (Cited on page 6.)
[62] B. Fredriksson. Finite element solution of surface nonlinearities in structural mechanics
with special emphasis to contact and fracture mechanics problems. Computers & Structures, 6(4–5):281–290, 1976. (Cited on page 6.)
[63] T. D. Sachdeva and C. V. Ramakrishnan. A finite element solution for the two–dimensional
elastic contact problems with friction. Int. J. Numer. Meth. Engng., 17:1257–1271, 1981.
(Cited on page 6.)
[64] L. T. Campos, J. T. Oden, and N. Kikuchi. A numerical analysis of a class of contact
problems with friction in elastostatics. Computer Methods in Applied Mechanics and Engineering, 34(1–3):821–845, 1982. (Cited on page 6.)
[65] J. A. C. Martins and J. T. Oden. A numerical analysis of a class of problems in elastodynamics with friction. Computer Methods in Applied Mechanics and Engineering, 40(3):327–360,
1983. (Cited on page 6.)
[66] N. Okamoto and M. Nakazawa. Finite element incremental contact analysis with various
frictional conditions. Int. J. Numer. Meth. Engng., 14:337–357, 1979. (Cited on page 6.)
[67] I. Babu˘ska. The finite element method with penalty. Math. Comp., 27:221–228, 1973.
(Cited on page 6.)
[68] N. Kikuchi. Convergence of a penalty–finite element approximation for an obstacle problem.
Numer. Math., 37(1):105–120, 1981. (Cited on page 6.)
[69] N. Kikuchi and J. T. Oden. Contact problems in elasticity: a study of variational inequalities and finite element methods, volume 8 of SIAM Studies in Applied Mathematics.
Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA, 1988. (Cited
on page 6.)
[70] N. D. Hung and G. Saxce. Frictionless contact of elastic bodies by finite element method
and mathematical programming technique. Computers & Structures, 11(12):55–67, 1980.
(Cited on page 6.)
[71] A. Klarbring and G. Bjorkman. A mathematical programming approach to contact problems with friction and varying contact surface. Computers & Structures, 30(5):1185–1198,
1988. (Cited on page 6.)
139
Bibliography
[72] J. C. Simo and T. A. Laursen. An augmented Lagrangian treatment of contact problems
involving friction. Comput. & Structures, 42(1):97–116, 1992. (Cited on page 6.)
[73] J. D. Lambert. Computational methods in ordinary differential equations. John Wiley
& Sons, London–New York–Sydney, 1973. Introductory Mathematics for Scientists and
Engineers. (Cited on page 6.)
[74] N. M. Newmark. A method of computational for structural dynamics. ASCE, Journal
of the Engineering Mechanics Division, Vol. 85 No. EM3,, 1959. (Cited on pages 6, 64
and 102.)
[75] H. M. Hilber, T . J. R. Hughes, and R. L Taylor. Improved numerical dissipation for
time integration algorithms in structural dynamics. Earthquake Engineering & Structural
Dynamics, 5:283–292, 1977. (Cited on pages 6 and 64.)
[76] P. Hauret and P. Le Tallec. Energy–controlling time integration methods for nonlinear
elastodynamics and low–velocity impact. Comput. Methods Appl. Mech. Engrg., 195(37–
40):4890–4916, 2006. (Cited on page 6.)
[77] N. J. Carpenter, R. L. Taylor, and M. G. Katona. Lagrange constraints for transient finite
element surface contact. Int. J. Numer. Meth. Engng., 32:103–128, 1991. (Cited on pages 6
and 56.)
[78] T. A. Laursen and V. Chawla. Design of energy conserving algorithms for frictionless
dynamic contact problems. Internat. J. Numer. Methods Engrg., 40(5):863–886, 1997.
(Cited on pages 6, 33 and 56.)
[79] V. Chawla and T. A. Laursen. Energy consistent algorithms for frictional contact problems.
Internat. J. Numer. Methods Engrg., 42(5):799–827, 1998. (Cited on pages 6 and 56.)
[80] T. A. Laursen and G. R. Love. Improved implicit integrators for transient impact problems–
geometric admissibility within the conserving framework. Internat. J. Numer. Methods
Engrg., 53(2):245–274, 2002. (Cited on pages 6, 33 and 56.)
[81] M. Barboteu. An energy–conserving algorithm for nonlinear elastodynamic contact problems – extension to a frictional dissipation phenomenon. Analysis and Simulation of Contact Problems, Lecture Notes in Applied and Computational Mechanics, 27:71–78, 2006.
(Cited on page 6.)
[82] K. Lee. A numerical solution for dynamic contact problems satisfying the velocity and
acceleration compatibilities on the contact surface. Computational Mechanics, 15(3):189–
200, 1994. (Cited on page 6.)
[83] F. Armero and E. Petocz. Formulation and analysis of conserving algorithms for frictionless
dynamic contact/impact problems. Comput. Methods Appl. Mech. Engrg., 158(3-4):269–
300, 1998. (Cited on page 6.)
140
Bibliography
[84] H. Khenous, P. Laborde, and Y. Renard. Mass redistribution method for finite element
contact problems in elastodynamics. Eur. J. Mech. A Solids, 27(5):918–932, 2008. (Cited
on pages 6, 7, 25, 33, 39, 56, 95, 109 and 112.)
[85] L. Paoli. Time discretization of vibro–impact. R. Soc. Lond. Philos. Trans. Ser. A Math.
Phys. Eng. Sci., 359(1789):2405–2428, 2001. Non–smooth mechanics. (Cited on pages 7,
33, 65, 106 and 112.)
[86] D. Doyen and A. Ern. Convergence of a space semi–discrete modified mass method for
the dynamic Signorini problem. Commun. Math. Sci., 7(4):1063–1072, 2009. (Cited on
page 7.)
[87] Y. Renard and J. Pommier. Getfem++. An Open Source generic C++ library for finite
element methods,. http://home.gna.org/getfem. (Cited on pages 9, 40, 65, 102 and 114.)
[88] Farshid Dabaghi, Adrien Petrov, Jérôme Pousin, and Yves Renard. Convergence of mass
redistribution method for the one-dimensional wave equation with a unilateral constraint
at the boundary. ESAIM: Mathematical Modelling and Numerical Analysis, 48:1147–1169,
7 2014. (Cited on pages 23, 56, 62, 63, 94, 95, 98 and 112.)
[89] M. Schatzman and M. Bercovier. Numerical approximation of a wave equation with unilateral constraints. Math. Comp., 53(187):55–79, 1989. (Cited on pages 25, 28, 36, 37, 54,
98 and 99.)
[90] C. Hager and B. I. Wohlmuth. Analysis of a space–time discretization for dynamic elasticity
problems based on mass-free surface elements. SIAM J. Numer. Anal., 47(3):1863–1885,
2009. (Cited on pages 25 and 63.)
[91] H. Brézis. Opérateurs maximaux monotones et semi–groupes de contractions dans les espaces de Hilbert. North-Holland Publishing Co., Amsterdam, 1973. North–Holland Mathematics Studies, No. 5. Notas de Matemática (50). (Cited on pages 26, 43, 59, 61 and 62.)
[92] W. Rudin. Real and complex analysis. McGraw-Hill Book Co., New York, second edition,
1974. McGraw-Hill Series in Higher Mathematics. (Cited on page 27.)
[93] J. Simon. Compact sets in the space Lp (0, T ; B). Ann. Mat. Pura Applic, pages 65–96,
1987. (Cited on page 27.)
[94] J. J. Moreau. Liaisons unilatérales sans frottement et chocs inélastiques. C. R. Acad. Sci.
Paris Sér. II Méc. Phys. Chim. Sci. Univers Sci. Terre, 296(19):1473–1476, 1983. (Cited
on pages 33 and 112.)
[95] J. J. Moreau and P. D. Panagiotopoulos. Nonsmooth mechanics and applications, volume
302 of CISM Courses and Lectures. Springer–Verlag, Vienna, 1988. (Cited on page 33.)
[96] L. Paoli and M. Schatzman. Schéma numérique pour un modèle de vibrations avec contraintes unilatérales et perte d’énergie aux impacts, en dimension finie. C. R. Acad. Sci.
Paris Sér. I Math., 317(2):211–215, 1993. (Cited on page 33.)
141
Bibliography
[97] L. Paoli and M. Schatzman. Approximation et existence en vibro–impact. C. R. Acad.
Sci. Paris Sér. I Math., 329(12):1103–1107, 1999. (Cited on page 33.)
[98] K. Schweizerhof and J. O. Hallquist. Efficiency refinements of contact strategies and algorithms in explicit finite element programming. Computational Plasticity, eds. Owen,
Onate, Hinton, Pineridge Pres, pages 457–482, 1992. (Cited on page 33.)
[99] M. Crouzeix and A. L. Mignot. Analyse numérique des équations différentielles. Collection
Mathématiques Appliquées pour la Maîtrise. [Collection of Applied Mathematics for the
Master’s Degree]. Masson, Paris, 1984. (Cited on pages 35 and 96.)
[100] A. Ern and J. L. Guermond. Theory and practice of finite elements, volume 159 of Applied
Mathematical Sciences. Springer–Verlag, New York, 2004. (Cited on page 38.)
[101] Y. Renard. Generalized Newton’s methods for the approximation and resolution of frictional contact problems in elasticity. Comput. Methods Appl. Mech. Engrg., 256:38–55,
2013. (Cited on page 40.)
[102] A. Curnier and P. Alart. A generalized Newton method for contact problems with friction.
J. Méc. Théor. Appl., 7(suppl. 1):67–82, 1988. (Cited on page 40.)
[103] P. Wriggers. Computational contact mechanics. John Wiley & Sons Ltd., 2002. (Cited on
pages 42 and 65.)
[104] K. Deimling. Multivalued differential equations, volume 1 of de Gruyter Series in Nonlinear
Analysis and Applications. Walter de Gruyter & Co., Berlin, 1992. (Cited on pages 43, 47,
48, 49 and 50.)
[105] J. M. Ball. Strongly continuous semigroups, weak solutions, and the variation of constants
formula. Proc. Amer. Math. Soc., 63(2):370–373, 1977. (Cited on pages 44 and 53.)
[106] J.P. Aubin. Approximation of elliptic boundary–value problems. Wiley–Interscience [A
division of John Wiley & Sons, Inc.], New York–London–Sydney, 1972. Pure and Applied
Mathematics, Vol. XXVI. (Cited on page 46.)
[107] V. Barbu. Nonlinear semigroups and differential equations in Banach spaces. Editura
Academiei Republicii Socialiste România, Bucharest, 1976. Translated from the Romanian.
(Cited on page 47.)
[108] K. J. Engel and R. Nagel. One–parameter semigroups for linear evolution equations, volume 194 of Graduate Texts in Mathematics. Springer–Verlag, New York, 2000. With
contributions by S. Brendle, M. Campiti, T. Hahn, G. Metafune, G. Nickel, D. Pallara, C.
Perazzoli, A. Rhandi, S. Romanelli and R. Schnaubelt. (Cited on page 51.)
[109] A. Pazy. Semigroups of linear operators and applications to partial differential equations,
volume 44 of Applied Mathematical Sciences. Springer–Verlag, New York, 1983. (Cited on
page 51.)
142
Bibliography
[110] T. Kato. Perturbation theory for linear operators. Springer–Verlag, Berlin, second edition,
1976. Grundlehren der Mathematischen Wissenschaften, Band 132. (Cited on page 52.)
[111] F. Dabaghi, A. Petrov, J. Pousin, and Y. Renard. Numerical approximations of elastodynamic contact problems based on mass redistribution method. Preprint, 2013. (Cited on
pages 55, 101, 109, 112, 113, 114 and 123.)
[112] T. A. Laursen. Computational contact and impact mechanics. Fundamentals of modeling interfacial phenomena in nonlinear finite element analysis. Springer–Verlag, Berlin,
Heidelberg, New York,, 2003. (Cited on pages 56, 64, 104 and 112.)
[113] P. Deuflhard, R. Krause, and S. Ertel. A contact–stabilized Newmark method for dynamical
contact problems. Internat. J. Numer. Methods Engrg., 73(9):1274–1290, 2008. (Cited on
page 56.)
[114] C. Hager, S. Hüeber, and B. I. Wohlmuth. A stable energy–conserving approach for frictional contact problems based on quadrature formulas. Internat. J. Numer. Methods Engrg.,
73(2):205–225, 2008. (Cited on page 56.)
[115] Y. Renard. The singular dynamic method for constrained second order hyperbolic equations: application to dynamic contact problems. J. Comput. Appl. Math., 234(3):906–923,
2010. (Cited on page 56.)
[116] C. Pozzolini and M. Salaun. Some energy conservative schemes for vibro–impacts of a beam
on rigid obstacles. ESAIM Math. Model. Numer. Anal., 45(6):1163–1192, 2011. (Cited on
page 56.)
[117] A. Tkachuk, B. I. Wohlmuth, and M. Bischoff. Hybrid–mixed discretization of elasto–
dynamic contact problems using consistent singular mass matrices. Internat. J. Numer.
Methods Engrg., 94(5):473–493, 2013. (Cited on page 56.)
[118] J. L. Lions and E. Magenes. Problèmes aux limites non homogènes et applications. Vol. 1.
Travaux et Recherches Mathématiques, No. 17. Dunod, Paris, 1968. (Cited on page 58.)
[119] H. Brezis. Functional analysis, Sobolev spaces and partial differential equations. Universitext. Springer, New York, 2011. (Cited on page 58.)
[120] T. J. R. Hughes. The finite element method. Prentice Hall Inc., Englewood Cliffs, NJ,
1987. Linear static and dynamic finite element analysis, with the collaboration of Robert
M. Ferencz and Arthur M. Raefsky. (Cited on pages 64, 102, 104, 112 and 123.)
[121] S. Krenk. Energy conservation in Newmark based time integration algorithms. Comput.
Methods Appl. Mech. Engrg., 195(44–47):6110–6124, 2006. (Cited on pages 64, 102, 104,
112 and 123.)
[122] E. Grosu and I. Harari. Stability of semidiscrete formulations for elastodynamics at small
time steps. Finite Elem. Anal. Des., 43(6-7):533–542, 2007. (Cited on pages 64, 104
and 112.)
143
Bibliography
[123] L. Paoli and M. Schatzman. A numerical scheme for impact problems. I ,II. The multidimensional case. SIAM J. Numer. Anal., 40(2):734–768, 2002. (Cited on pages 65 and 106.)
[124] Y. Dumont and L. Paoli. Vibrations of a beam between obstacles. Convergence of a fully
discretized approximation. M2AN Math. Model. Numer. Anal., 40(4):705–734, 2006. (Cited
on pages 79 and 106.)
[125] F. Dabaghi, A. Petrov, J. Pousin, and Y. Renard. Numerical study of convergence of the
mass redistribution method for elastodynamic contact problems. Congress proceedings of
WCCM–ECCM–ECFD, 2014. (Cited on page 109.)
144
Etude de la convergence des méthodes de redistribution de masse pour les
problèmes de contact en élastodynamique
Résumé: Le chapitre 1 porte sur une équation des ondes monodimensionnelle soumise à une condition
aux limites unilatérale. Sous des hypothèses de régularité appropriées sur les données initiales, une
nouvelle preuve d’existence et d’unicité est proposée. La méthode de redistribution de masse qui repose
sur une redistribution de la masse d’un corps de telle sorte qu’il n’y ait pas d’inertie au niveau du nœud
de contact est introduite et sa convergence est prouvée. Une approximation de ce problème d’évolution
combinant la méthode des éléments finis ainsi que la méthode de redistribution de masse est analysée
dans le chapitre 2. Puis deux problèmes ainsi que leurs solutions analytiques respectives (l’une étant
nouvelle) sont présentés et des discrétisations possibles en utilisant différentes méthodes d’intégration en
temps sont décrites. Enfin, des simulations numériques de ces problèmes sont reportées. Dans le chapitre
3, la masse des nœuds de contact est redistribuée sur les autres nœuds et sa convergence ainsi qu’une
estimation de l’erreur en temps sont établies. Ensuite, une solution analytique déjà introduite dans le
chapitre 3 est comparée aux approximations obtenues pour plusieurs redistributions de masse possibles
mettant ainsi en évidence que plus une redistribution de masse d’un corps se fait à proximité des nœuds de
contact meilleures sont les solutions approchées obtenues. Les problèmes de contact élastodynamique en
dimension d’espace deux et trois sont étudiés dans le chapitre 4. Comme pour les problèmes de contact
monodimensionnels, une solution approchée combinant les éléments finis et la redistribution de masse
est exposée. Quelques simulations numériques utilisant des méthodes d’intégration en temps mettent en
évidence les propriétés de convergence de la méthode de redistribution de masse.
Mots clés: Elastodynamique, existence, unicité, méthode de redistribution de masse, inégalité variationnelle, contact unilatéral, conservation de l’énergie, solution numérique.
Study of the convergence of the mass redistribution method for the
elastodynamic contact problems
Abstract: The chapter 1 focuses on a one–dimensional wave equation being subjected to a unilateral
boundary condition. Under appropriate regularity assumptions on the initial data, a new proof of existence and uniqueness results is proposed. The mass redistribution method based on a redistribution of the
body mass such that there is no inertia at the contact node is introduced and its convergence is proved.
An approximation of this evolutionary problem combining the finite element method as well as the mass
redistribution method is analyzed in chapter 2. Then two benchmark problems (one being new) with
their analytical solutions are presented and some possible discretizations using different time–integration
schemes are described. Finally, numerical experiments for these benchmark problems are reported. In
chapter 3, the mass of the contact nodes is redistributed on the other nodes and its convergence as well
as an error estimate in time are established. Then an analytical solution already introduced in chapter
3 is compared to approximate ones obtained for different choices of mass redistribution highlighting that
more a mass redistribution of the body is done near the contact nodes better the approximate solutions
are obtained. The two and three–dimensional elastodynamic contact problems are studied in chapter 4.
As for the one–dimensional contact problems, an approximated solution combining the finite element and
mass redistribution methods is exhibited. Some numerical experiments using time–integration methods
highlighted the convergence properties of the mass redistribution method.
Keywords: Elastodynamics, existence, uniqueness, mass redistribution method, variational inequality,
unilateral contact, energy conservation, numerical solution.